Files
aevolution/docs/engineering/VALIDATION.md
T
dingoandClaude Opus 5 cabada70c4 Account for the interval a model proposes; never clamp it
M4.3 advances a cohort over one declared interval and deliberately does not model
one. ECOLOGY_MODEL keeps population maths an experiment and asks for several
candidates to be compared, so choosing one here would turn a recommendation into
a locked rule. A model proposes fluxes; this checks the declared result is
exactly what they imply, for counts and for every inventory channel.

Nothing is clamped. A loss larger than what is there refuses the interval and
says whether it was the population or the biomass, so a caller can subdivide;
subtraction is checked against what is present rather than performed and
repaired, so a negative population is never constructed. That is the answer to
the named risk: a clamp is how an interval a model could not integrate becomes a
plausible result, and how a false extinction arrives. Both clamp mutations are
detected by the tests.

Counts carry whole members and owed fractions as one quantity, so a population
below one member is a number rather than a rounding decision. Biomass has no
birth term at all, so a birth cannot bring inventory in and newborns are
necessarily funded from what the cohort holds. Immigration draws from a reservoir
that already held it, so a late transfer is not new production. Conservation runs
through the ledger as one atomic transaction; an interval where nothing moves
submits none.

Two amendments to M4.1 were forced, not chosen. A cohort records the reservoir
backing it, because an advance cannot otherwise know where the inventory is and
passing it alongside would let a caller advance one cohort against another's
store, which comparing balances cannot detect between equal reserves; the
registry now refuses two cohorts sharing one. And a cohort may be vacant, through
a separate factory so vacancy stays deliberate, because a decline to exactly zero
was otherwise unrepresentable and the only way through would have been the clamp
this milestone exists to refuse. A vacant cohort is not an extinction: the
record, its genome, its habitat and its provenance remain and it stays installed.

Fourteen guards confirmed load-bearing by removing each and rerunning. Two of my
own expectations were wrong rather than the code: escrow arithmetic in the
migration test, and a dwindling case that shed all but one femtogram while
declaring a result holding none. The second surfaced a real tension now recorded
in ADR-0038 rather than worked around: a class reduced to a fraction of a member
may hold no reserve, because reserve belongs to members and a fraction is not one
yet, so that biomass has to go somewhere the same interval.

Closure is checked and stability is not, so an interval whose arithmetic balances
and whose ecology is nonsense is accepted. The declared age structure is not
verified to be what ageing would produce, because that needs the within-class
distribution the record discards.

Release 36/36, analysis, ASan 36/36 and clang-format all ran and passed here.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
2026-09-24 10:23:09 -04:00

1137 lines
72 KiB
Markdown
Raw Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
# Validation report
Recorded 2026-09-19. Latest verified Linux delivery: **run 1402 passed** on
`4c2cc04ac47d37c6191a9ef02d4bbba716a930dc`. Native checks apply to base commit `8ab48d3` plus the
foundation-delivery working changes committed with this report. Simulation rules
and snapshot format remain version 1; the only C++ change is benchmark metadata.
Results below distinguish executed checks from pending Linux/cluster gates.
## Tools and host
- Windows NT 10.0.26200, AMD Ryzen 7 9800X3D, 8 cores / 16 logical processors.
- MSVC 19.44.35217.0, CMake 3.31.6-msvc6, Ninja 1.12.1.
- clang-format and clang-tidy 19.1.5 (Visual Studio LLVM tools).
- kubectl 1.31.4 / Kustomize 5.4.2 for offline manifest rendering.
## Executed native checks
| Command | Observed result |
|---|---|
| `./tools/windows-check.ps1 -Preset debug -Action check` | Configure/build passed; 5/5 CTests passed |
| `./tools/windows-check.ps1 -Preset release -Action check` | Configure/build passed; 5/5 CTests passed under NDEBUG |
| `./tools/windows-check.ps1 -Preset analysis -Action check` | clang-tidy/warnings-as-errors build passed; 5/5 CTests passed |
| `./tools/windows-check.ps1 -Preset asan -Action check` | MSVC AddressSanitizer build passed; 5/5 CTests passed; this is NOT UBSan |
| `./tools/windows-check.ps1 -Preset debug -Action format` | Passed with clang-format 19; repeated twice in one shell after helper fix |
| `./build/release/evo-sim --seed 123456 --entities 8 --ticks 1000000 --verify-replay` | Exit 0; replay_verified=true; hash fab22c2e8ae1c379 |
| `./build/release/evo-benchmark` | Exit 0; MSVC 1944 Release, seed 123456, 10000 entities, 1000 ticks, 0.0210666 seconds, hash 3bad6db1612cbd1 |
| Git Bash `bash -n tools/ci/linux-check.sh` and `bash -n tools/ci/publish-image.sh` | Syntax checks passed; no Linux execution implied |
| `kubectl kustomize deploy/kubernetes/dev` | Rendered Job, ServiceAccount and NetworkPolicy in evolution-dev; final render pins the verified run-1402 image |
CTest covers the independently frozen 1000-tick fingerprint d61ad5bf0c928c76,
seeded replay properties, malformed/semantic snapshot rejection, entity lifecycle,
capacity/overflow, telemetry and CLI file continuation/alias protection. The
expected-failure harness confirms tests remain active in Release.
The benchmark is a single uncontrolled wall-clock observation, not a performance
threshold or an MMO capacity estimate. Its initial compiler output accidentally
used hexadecimal; the stream now returns to decimal and reports MSVC 1944.
## Gitea evidence
The maintainer's [run 1399](https://git.aeternalisgames.com/dingo/AEvolution/actions/runs/1399)
on `8ab48d3a5639121693f5881b600482e870b49f05` completed successfully. API job records
confirm verify (configure/build/CTest) and publish (Docker build/push) succeeded
on runner theraborn-runners-runner-0, label ubuntu-22.04. This is evidence for the
original workflow, not the expanded sanitizer/analysis jobs added in this change.
Exact Linux tool versions and expanded-run results appear below; a runner label
is not a compiler version.
## Limitations and next gates
- Local Docker Desktop has no running Linux engine. Container checks must run on
the configured Gitea runner; no local container success is claimed.
- Expanded Linux GCC14/Clang19, ASan/UBSan, TSan and libFuzzer evidence is recorded
below. TSan currently validates a sequential kernel, not future concurrency.
- Sandbox process restrictions stalled native Ninja and blocked Git Bash signal
pipes. Those attempts were stopped; checks reran successfully outside sandbox.
- Repeated VsDevCmd initialization grew PATH beyond cmd's limit. The helper now
reuses the matching x64 environment and deduplicates PATH; repeated checks pass.
- The maintainer reports Argo setup, but this workspace has no Kubernetes context.
No dev Job completion or cluster admission result is recorded yet.
- No biology, network/load, durable server recovery or backup restore exists.
Those remain M2, M7/M8 and later gates, not failing or passing foundation tests.
Update this report with run URL, tested commit, image digest and Job result when
those checks execute. Never overwrite a golden merely to make CI pass.
## Expanded Linux run 1400
[Run 1400](https://git.aeternalisgames.com/dingo/AEvolution/actions/runs/1400),
commit 9e412736d91ad103b7a951f99b811c629c5ac61e: GCC14 Debug and Release, Clang19
analysis and ASan/UBSan each passed 5/5 tests. Formatting passed. The seeded fuzz
smoke completed 10000 cases with no reported failure. CMake 3.28.3, GCC 14.2.0,
Clang/format/tidy 19.1.1, Linux 6.8.0-110-generic x86-64. Million-tick hash matched
Windows (fab22c2e8ae1c379); benchmark hash matched too (3bad6db1612cbd1), with
0.0674131 seconds observed for 10000 entities * 1000 ticks on the CI host.
TSan compiled but all executions failed before main: its runtime could not call
personality(ADDR_NO_RANDOMIZE) inside the Docker build sandbox. The image job was
skipped, as required. This is not recorded as a TSan pass. The follow-up isolates
execution in a container with the narrowly constrained profile from ADR-0020;
its result must be recorded after execution.
## Successful delivery run 1402
[Run 1402](https://git.aeternalisgames.com/dingo/AEvolution/actions/runs/1402),
commit 4c2cc04ac47d37c6191a9ef02d4bbba716a930dc: every job succeeded. The general
checks reused Docker's valid cache for unchanged checked sources; their full
execution also passed in run 1401. TSan executed all five CTests successfully in
2.71 seconds using the scoped profile and writable tmpfs outputs. Run 1401 had
first exposed a non-root tmpfs permission error; setting explicit mode=1777 on
these isolated scratch mounts fixed it. No test was disabled or ignored.
The runtime image built, ran the restricted replay smoke test and was pushed to:
`registry.aeternalisgames.com/aeternalis/aevolution:sha-4c2cc04ac47d37c6191a9ef02d4bbba716a930dc-run-6-1`
Registry digest:
`sha256:8fce2b462c9907f38d3fb832afeb70d0bff068dd832b2171e861186dbcb8e6b9`.
The dev manifest pins this digest. This proves build/test/image delivery; Argo
application mapping and Kubernetes Job completion remain unverified locally.
## M1.5 native ECS experiment
Opt-in target evo-ecs-experiment (EnTT4 pinned in ADR-0021), based on 1c6e6e1 plus
this experiment's changes. MSVC Release contract checks passed, including 72
seed/density/count combinations, hand-calculated arithmetic, canonical equality,
restore continuation, retired-ID rejection, malformed snapshot checks and
allocator balance. The final clang-tidy build/test passed after tightening the
process-boundary diagnostic catch. MSVC ASan passed on the final source.
Baseline Release CTest remained 5/5
and format-check included the experiment. No simulation rule/golden changed.
Commands: configure release/analysis/asan with `-B build/ecs[-analysis|-asan]
-DEVO_BUILD_ECS_EXPERIMENT=ON`, build target evo-ecs-experiment, then
`ctest --test-dir <directory> -R ecs-experiment --output-on-failure`.
Seven measured samples plus warmup are retained in results/ecs-msvc1944.jsonl;
protocol, build observations and scope limits are in ECS_EVALUATION.md. The
Linux experiment CI job is pending its first pushed run; no Linux experiment
success is claimed yet.
## M1.5 Linux evidence
[Run 1404](https://git.aeternalisgames.com/dingo/AEvolution/actions/runs/1404),
source bf7d60b7cd174736d38aea42616064e55e0c5f05: every job passed. The new ci-ecs
stage explicitly prepared the pinned header/license, then built and ran the
experiment under GCC14 Release, Clang19 ASan/UBSan and Clang19 analysis. All three
contract invocations passed. Raw measured GCC samples are retained at
results/ecs-gcc142.jsonl; all 56 fingerprints match corresponding Windows samples.
No production rule or snapshot-format change occurred. This supersedes the
pending Linux experiment note above. The final follow-up edits only record
results and decision scope; production ECS adoption remains deferred.
## M1.6 sequential scheduler (2026-09-19)
Native commands: `tools/windows-check.ps1 -Preset release -Action configure`,
clang-format on changed C++ files, `cmake --build --preset release`,
`ctest --preset release`; then `tools/windows-check.ps1 -Preset analysis -Action check`,
`tools/windows-check.ps1 -Preset asan -Action check` and Release `-Action format`.
All three configurations passed 5/5 CTests; clang-tidy and format-check passed.
Tests execute under NDEBUG and cover six batch permutations across 32 seeds,
an independent direct-lifecycle reference, ordered receipts, restore continuation,
rejected conflicts and malformed admission, the 1,024-command boundary, full entity
capacity and tick/ID exhaustion. Golden `d61ad5bf0c928c76` remains unchanged.
The copy-and-commit implementation allocates before publishing world state and
statically checks noexcept publication/receipt moves. Allocation-failure injection
was not run. No scheduler throughput claim, concurrency validation, durable input
journal or Kubernetes execution is implied.
[Gitea run 1407](https://git.aeternalisgames.com/dingo/AEvolution/actions/runs/1407)
passed all four jobs for source `b620edfd6f8097172acf055402a24620cb078358`:
Linux checks, the ECS experiment, separate TSan, and image smoke/publication.
This closes the pending Linux validation for M1.6. The maintainer's independent
dev image pin from `20e609d` was preserved during rebase. Kustomize renders the
three expected dev resources successfully; no cluster execution is claimed.
## M1.7 spatial queries (2026-09-19)
Release and MSVC ASan passed 5/5 CTests. Tests cover periodic-image reference
comparison (32 seeds × 3 lifecycle/restore phases × 24 queries), hand-checked
seams/corners/inclusive radii, empty/full capacity, invalid centers and maximum
u32 radius. Queries leave snapshots unchanged; existing v1 goldens remain intact.
Native commands use `tools/windows-check.ps1 -Preset <release|analysis|asan>
-Action check` and Release `-Action format`. Initial analysis flagged an unnecessary
test-helper vector copy; the helper now takes a const reference. Final analysis,
Release, ASan and formatting passed after applying the helper's required line
wrapping. Linux validation is recorded below.
`build/release/evo-spatial-benchmark.exe` produced 72 measured records plus
metadata in results/spatial-msvc1944.jsonl. One warmup and three samples per
scenario; 1k/10k/100k entities, 128 queries, uniform/clustered positions, four radii.
Timing includes result allocation, destruction and checksum work, excludes setup;
no performance threshold. Desktop CPU load was not isolated. No cluster changes
are required for this library increment.
[Gitea run 1409](https://git.aeternalisgames.com/dingo/AEvolution/actions/runs/1409)
passed all four jobs for source `9b19509673d88683ee72e093e42534bfe151f928`:
Linux checks, ECS regression, separate TSan and image smoke/publication.
All 72 spatial benchmark checksums and match counts agree with Windows.
Raw Linux measurements are in results/spatial-gcc142.jsonl; 24 native scenarios
also have identical checksums across their three repetitions. No deployment
manifest, rules version, snapshot format or golden was changed.
## M2.1 finite inventory ledger (2026-09-19)
Native Release, analysis and MSVC ASan each passed 6/6 CTests; formatting passed.
Commands: `tools/windows-check.ps1 -Preset <release|analysis|asan> -Action check`
and Release `-Action format`. Initial clang-tidy rejected the Element enum's
oversized base type; changed to u8 and reran all configurations successfully.
The new independent ledger test checks typed energy/mass separation, finite-source
exhaustion, explicit heat transfer, intermediate overdraft, whole-transaction
rollback, retry/duplicate action semantics, invalid reservoir IDs, maximum batch
and reservoir counts, opening-total overflow and exact u64-limit transfers.
32 seeded runs × 31,250 transactions = one million transactions, each checked
against independent reference balances, a replay branch and recomputed per-channel
totals. This is transfer replay, not serialized recovery or an ecosystem run.
Final ledger tests took 0.63s Release and 7.62s ASan locally; not controlled
performance measurements. Existing particle goldens remain unchanged.
Allocation-failure injection, durable ledger recovery and cross-ledger ownership
are not tested or implemented. Rates and conversions are absent, so rounding
residuals cannot occur in this transfer-only scope.
[Gitea run 1411](https://git.aeternalisgames.com/dingo/AEvolution/actions/runs/1411),
source `8fd3591b0e37aaecf00a6410a5926745a031cb6f`, passed
the Linux checks, ECS regression and separate TSan jobs. Each main Linux
configuration passed 6/6 CTests. Ledger-suite times were 3.34s GCC Debug, 0.37s GCC
Release, 2.46s Clang analysis and 8.78s Clang ASan/UBSan; these are CI test
durations, not controlled performance benchmarks. The image smoke/publication
job also passed, completing all four jobs. No deployment image pin was changed.
## M2.2 finite producer process (2026-09-19)
Native Release, clang-tidy and MSVC ASan passed 7/7 CTests; format-check passed.
Commands: `tools/windows-check.ps1 -Preset <release|analysis|asan> -Action check`
and Release `-Action format`. The producer suite compares hand-calculated ticks
and 32 seeds × 1,000 ticks against a small-domain cumulative arithmetic reference,
an independent replay branch and per-tick energy/element conservation audits.
It covers invalid config/ticks with unchanged state, fractional capture/growth,
efficiency endpoints, finite source exhaustion, all four limiting elements,
zero capture, maximum u64 capture with 50% carried conversion, and maximum recipe
and growth-cost multiplication. Particle goldens remain unchanged.
Allocation-failure injection and tick-counter exhaustion were not exercised;
the latter has an explicit guard but no arbitrary-state restore API. Tests use
in-memory copies, not serialized recovery. No photosynthetic chemistry, metabolic
model, database integration, throughput guarantee or deployment is claimed.
[Gitea run 1413](https://git.aeternalisgames.com/dingo/AEvolution/actions/runs/1413)
passed all four jobs for source `af293b8b50cbb506a49dfff10c6cc92307e16aca`:
Linux checks, ECS regression, separate TSan and runtime image smoke/publication.
The initial push was delayed by a tool approval-review usage limit; retry succeeded
after the maintainer resumed work. No deployment image pin was changed.
## M2.3 shared feeding community (2026-09-20)
Native Release, clang-tidy and MSVC ASan each passed 8/8 CTests; formatting passed.
Commands: `tools/windows-check.ps1 -Preset <release|analysis|asan> -Action check`
and Release `-Action format`. The new suite checks hand-calculated capture/feeding,
producer-to-consumer and consumer-to-consumer predation, heat/waste outputs,
terminal actor rejection, full-tick rollback, competing resource requests,
duplicate sequence/consumer admission, per-channel intake caps, invalid actors,
config/capacity limits and u64 assimilation endpoints. 32 seeds × 200 ticks compare
reordered input batches, receipts, full balances and producer progress, with
independently recomputed energy and elemental totals. Standalone producer and
existing particle golden tests remain unchanged and passing.
No serialized recovery, allocation-failure injection, spatial reach/perception,
physical injury, corpse handling or persistent-world integration was tested or
implemented. These are co-located synthetic process fixtures.
[Gitea run 1415](https://git.aeternalisgames.com/dingo/AEvolution/actions/runs/1415)
passed all four jobs for source `b7dd466f57f385d6a8ffccfaa69f639077545587`:
Linux checks, ECS regression, separate TSan and image smoke/publication.
No deployment image pin was changed.
## M2.4 terminal inventory and decomposition (2026-09-20)
Native Release, analysis and MSVC ASan passed 8/8 CTests; formatting passed.
Commands: `tools/windows-check.ps1 -Preset <release|analysis|asan> -Action check`
and Release `-Action format`. Tests preserve corpse inventories on death, suppress
producer activity immediately, reject repeated/invalid terminal events and roll
back death with later invalid feeding. Sixteen seeds × 256 ticks compare a
closed-form depletion reference, reversed terminal-event order, per-tick budgets,
inventories and conservation. Additional cases cover shared corpse/waste limits,
mixed and zero channel caps, u64-maximum inventories, diagnostic rollback, and
returned nutrients funding producer structure on the next tick.
The existing feeding, standalone producer, ledger and particle tests pass unchanged.
This does not validate reaction chemistry, actual death causes, corpse scavenging,
serialized recovery, allocation-failure injection or persistent history. There is
no deployment change.
[Gitea run 1417](https://git.aeternalisgames.com/dingo/AEvolution/actions/runs/1417)
passed all four jobs for source `e17a9a7279d6bde736261c0c33abe46108e1f622`:
Linux checks, ECS regression, separate TSan and image smoke/publication.
## M2.5 funded reproduction (2026-09-21)
Native Release, analysis and MSVC ASan passed 8/8 CTests; formatting passed.
Commands: `tools/windows-check.ps1 -Preset <release|analysis|asan> -Action check`
and debug `-Action format`. The first format run rejected three wrapped test
declarations; clang-format 19.1.5 rewrote them and Release was rebuilt and rerun.
Tests cover hand-calculated endowment and construction overhead, the endowment
floor, per-channel insufficiency with abundant energy, zero and oversized
investments, reserved slots rejected as feeding targets, corpses and ancestry
subjects, the producer rejected as a parent, terminated and unknown parents,
zero/duplicate sequences, two intents naming one parent, oversized batches,
capacity exhaustion rejecting the whole tick, absent identity reuse after a child
dies and decomposes, ancestry surviving both deaths, and construction endpoints of
0 and 1000 including a paid-for zero-endowment child. 32 seeds × 4 ticks × 2 births
compare reversed intent order against balances, producer progress, birth receipts
and independently recomputed conservation. A u64-maximum investment endows exactly
and conserves. Default zero capacity leaves the M2.4 reservoir layout, feeding,
producer, ledger and particle tests unchanged; golden `d61ad5bf0c928c76` is intact.
Review of `c84a701` found one defect (2026-09-22): births validated each parent
against the progressively modified candidate, so a child allocated earlier in the
same batch was already active and could fund a second birth. Two generations could
therefore appear in one tick, which ADR-0028 forbids. Eligibility is now fixed on
entry to the reproduction phase, after terminal facts and feeding, and checked for
every intent before any birth is applied. The contract itself was already correct
and did not change. The regression test admits one intent whose parent is created
by the preceding intent and requires `unborn_actor` with the entire tick unchanged;
it was confirmed to fail against the unguarded implementation and to pass with the
guard. The same pair succeeds across two ticks, recording the grandchild ancestry.
All four native configurations were rerun after the fix.
Allocation-failure injection, serialized recovery of states/parents/origins and
long-run ecosystem behavior were not exercised. This fixture defines no
reproductive strategy, fecundity, mutation, gestation, spatial placement or account
ownership, and no deployment manifest or image pin changed. Linux validation is
pending its first pushed run; no Gitea result is claimed for this increment yet.
## M2.6 measured ecosystem episodes (2026-09-22)
Native Release, analysis and MSVC ASan each passed 11/11 CTests; formatting passed.
Commands: `tools/windows-check.ps1 -Preset <release|analysis|asan> -Action check`
and debug `-Action format`. Formatting rejected 25 wrapped declarations on the
first run; clang-format 19.1.5 rewrote them and Release was rebuilt and rerun.
Three CTest cases are new: `ecosystem-contracts`, `cli-ecosystem` pinned to
fingerprint `2cd4a15f9fbe9949`, and `cli-ecosystem-conflict`, which requires the
CLI to reject `--ecosystem` combined with proof-run options.
The suite checks a solitary founder that only ages, a hand-checked predation
retirement, births exhausting exactly the reserved slots, tick-limit termination
with survivors, configuration rejection, and two episodes interleaved without
observing each other. Across 64 seeds it steps tick by tick, requiring that the
driver never provokes a rejected batch and that conserved totals are identical at
every tick, then replays each seed and requires an identical report. That ensemble
advanced 22,804 ticks with 879 births, 730 predation deaths, 277 lifespan deaths
and extinction in all 64 episodes, episode length 73 to 463 ticks.
`build/release/evo-sim.exe --ecosystem --seed 0 --episodes 3000` advanced
**1,093,007 ticks** in 1.34 s, with 41,138 births, 33,984 predation deaths, 13,154
lifespan deaths, extinction in all 3,000 episodes, episode length 50 to 497 ticks
and fingerprint `130d7725c0dc8b16`. Deaths equal founders plus births exactly
(6,000 + 41,138 = 47,138), so every process died once and no producer was grazed
out: the producer always retains captured energy, so its whole inventory never
reaches zero. The elapsed time is one uncontrolled wall-clock observation, not a
performance threshold.
Review of `eaf58f2` found one defect (2026-09-22): the CLI validated mode options
in one direction only. `--ticks` was accepted and ignored under `--ecosystem`, and
`--episodes`/`--tick-limit` were accepted and ignored without it, so a measurement
script could exit zero while running the other workload. The parser now records
which options were supplied and rejects any that do not belong to the selected
mode, in both directions. Six CLI cases cover this; `cli-ecosystem-rejects-ticks`
and `cli-episodes-require-ecosystem` are the two genuine regressions, while the
entities, save and verify-replay cases were already rejected and add coverage.
The remaining shared options were audited: `--seed` and `--output` are valid in
both modes, and `--load` conflicts were already enforced. CTest is now 16 cases,
all four native configurations were rerun, and `2cd4a15f9fbe9949` is unchanged.
This measures arithmetic conservation and turnover, not ecological plausibility.
Strategy constants, lifespans and the targeting rule are chosen inputs. There is
no maintenance cost, starvation rule, perception, spatial gate, genome or
heritable variation, so lineages differ only by founder strategy. Episode state is
not serialized, allocation failure is not injected, and no population abstraction
exists. Golden `d61ad5bf0c928c76` and every earlier suite are unchanged. No
deployment manifest or image pin changed. Linux validation is pending its first
pushed run; no Gitea result or cross-platform fingerprint agreement is claimed yet.
## M3.1 immutable genome record (2026-09-22)
Native Release, analysis and MSVC ASan each passed 17/17 CTests; formatting
passed. Commands: `tools/windows-check.ps1 -Preset <release|analysis|asan>
-Action check` and debug `-Action format`. Formatting rejected three wrapped
declarations on the first run; clang-format 19.1.5 rewrote them and Release was
rebuilt and rerun. `genome-contracts` is the one new CTest case, covering a new
`evo::biology` library that nothing else links.
Tests cover founder and descendant records, unordered drafts producing byte-equal
canonical records, digests differing for identical capabilities from different
origins, and every ancestry rule: zero identity, zero origin, zero development
version, a zero parent, a self parent, two parents and a repeated parent. Schema
rules cover an empty record, each of the three required capabilities removed in
turn, a duplicate kind, an unknown kind, an unknown prerequisite bit, and both
inclusive boundaries of all seven declared ranges accepted with one below and one
above each rejected. Graph rules cover a declared prerequisite naming an absent
capability, self, mutual and three-node cycles, a valid ordering using the same
edges without the closing one, and a cycle closed by a mandated rather than a
declared edge.
Encoding tests cover the round trip, byte-exact re-encoding, truncation at every
length from zero to one byte short, one trailing byte, an oversized buffer, a
corrupted magic, and every single-byte corruption of the full record. Three cases
reseal the checksum after editing so the later checks are reached rather than
masked: a rewritten genome ID is caught by the digest, a future schema version
fails closed, and non-canonical record order is caught as a duplicate kind. A
parent count beyond the reserved room and a capability count beyond the schema
are both refused before the corresponding region is read.
Coverage limitation: in schema 1 every mandated capability is also a required
one, so the presence half of the mandate check is reachable only through declared
prerequisites, which is tested. The mandated edges do take part in the cycle
closure, which is tested. A later schema that mandates a capability without
requiring it would exercise the remaining branch; this is recorded rather than
covered by a test that could not fail.
Review of `8584eb1` found two defects (2026-09-22). First, `Genome` had implicit
move operations, so moving one emptied the source's capabilities while leaving
its identity, origin and digest intact: a record that still reported ID 7 but
whose encoded bytes no longer decoded. That contradicted the documented invariant
directly. The type is now copy-only, which suppresses the implicit moves so a
move decays to a copy, and the invariant holds for every operation. The new move
tests were confirmed to fail without the fix: the Debug and ASan builds exit 1 at
runtime, and the Release build does not compile at all, because MSVC proves the
moved-from comparison always throws and rejects the unreachable code under
warnings-as-errors. clang-tidy then flagged the deliberate decaying moves in the
tests as `performance-move-const-arg`; those two lines carry a NOLINT explaining
that the decay is the contract under test.
Second, the encoding tests compared the current encoder against the current
decoder, so a coordinated change to layout or digest would have passed while
breaking existing v1 records. A frozen 83-byte fixture now holds one schema 1
founder record. Every structural byte was verified by hand against
schemas/genome-v1.md before being frozen: magic, both versions, ID 7, origin 1,
zero parents, three records, kinds 0/1/5 with parameters 400/120/5000, and the
`56 + 9N` length. The digest `9b2cdedef16ac87f` and the trailing checksum are
computed values frozen alongside them. The fixture is asserted in both
directions: the encoder must still produce those bytes, and the decoder must
still read them into the expected record. It is a golden and must not be edited
to make a test pass.
No genome affects any simulation, and nothing links the new library except its
own tests. The M2 process fixtures still duplicate founder strategies and do not
read the record. Development, phenotype derivation, mutation, recombination,
multi-parent heredity, migration between schema versions and any persistent
genome archive are not implemented. The encoding is a local record format, not a
network protocol or durable storage, and neither the digest nor the checksum is
an authenticity claim. Golden `d61ad5bf0c928c76` and the ecosystem fingerprint
`2cd4a15f9fbe9949` are unchanged. No deployment manifest or image pin changed.
Linux validation is pending its first pushed run; no Gitea result is claimed yet.
## M3.2 development and phenotype (2026-09-22)
Native Release, analysis and MSVC ASan each passed 18/18 CTests; formatting
passed. Commands: `tools/windows-check.ps1 -Preset <release|analysis|asan>
-Action check` and debug `-Action format`. `development-contracts` is the one new
CTest case, in the existing `evo::biology` library, which still nothing else
links. No existing suite changed.
The fixture genome uses a target size of 8,000,000 fg, which is an exact 20
micrometre cube at the declared density, so every derived value is hand-calculable
and is asserted as a literal: structural mass 8,000,000, extent 20, surface 2,400,
membrane mass 28,800,000, total 36,800,000 and maintenance 160,000 picojoules per
tick. Half-built, damaged and maximum-size cases are hand-calculated the same way,
including the 1e9 fg body that is an exact 100 micrometre cube with an uptake rate
of exactly the reference surface, and the 999,999,999 fg body one cubic micrometre
short of it, whose extent is 99.
Determinism is checked over 45 growth and temperature combinations, each derived
twice and compared, with more growth never yielding less structure. Damage is
checked to change the phenotype while leaving the genome's encoded bytes and
content digest identical, which is the M3.2 risk stated in the roadmap. Capability
prerequisites are checked both ways: without the capability a component is zero at
any size, and with it uptake scales with surface while movement speed does not.
Thermal shortfall is checked inside tolerance, symmetrically outside it, and
without the capability, and is confirmed to change no other derived value.
Unsupported rule versions and out-of-range growth fail closed through both entry
points.
Plans are checked for the full funded step, for the exact largest affordable step
one femtogram short of it, for no growth with no material, for a finished organism,
for clamping the last step to the target rather than overshooting, and for never
repairing damage. The declared flooring is pinned rather than hidden: a body whose
mass floors to zero advances its growth counter for no material, which grants no
structure because none exists at that size either way.
Coverage limitation: the overflow guards are unreachable through the public API at
schema 1 ranges, because the largest admissible genome produces a total mass of
about 7.2e11 femtograms, eleven orders of magnitude below the u64 limit. They are
kept as guards for later schema ranges and are recorded here rather than covered
by a test that could not fail.
No organism develops, nothing links the library except its own tests, and no
inventory moves: a plan reserves nothing and pays for nothing. The transfer half
of ApplyDevelopment, energetic budgets, repair, ageing, starvation, mutation and
morphology are not implemented. Golden `d61ad5bf0c928c76`, ecosystem fingerprint
`2cd4a15f9fbe9949` and the frozen genome fixture are unchanged. No deployment
manifest or image pin changed. Linux validation is pending its first pushed run.
## M3.3 metabolic budgets (2026-09-23)
Native Release, analysis and MSVC ASan each passed 19/19 CTests; formatting
passed. Commands: `tools/windows-check.ps1 -Preset <release|analysis|asan>
-Action check` and debug `-Action format`. `metabolism-contracts` is the one new
CTest case. `evo_biology` now links `evo::accounting` and `evo_processes` links
`evo::biology`; no existing suite changed and no existing fixture reads genomes.
Every budget is asserted as a hand-calculated literal against the same 20
micrometre cube fixture as M3.2: body 24,160,000 / 2,640,000 / 3,280,000 /
6,720,000 femtograms across C/N/P/other, maintenance 160,000 and movement 184,000
picojoules per tick, intake 240,000 picojoules and 2400/720/240/1440 femtograms,
and a reproduction investment of 11,040,000 femtograms priced at 441,600
picojoules. Construction is checked to compose each compartment separately rather
than by one blended ratio, and to agree exactly with what `plan_development`
reports as material and energy.
The non-uniformity the increment exists to demonstrate is tested directly.
Doubling membrane thickness raises body matter to 65,600,000 femtograms,
maintenance to 304,000 picojoules and the movement cost, shifts the carbon share
upward because membrane is the carbon-rich compartment, and leaves intake limit,
speed, assimilation and structural mass identical; the mass and maintenance ratios
are asserted to differ, so no single multiplier could produce this. Growth is
tested the same way: uptake follows surface while maintenance follows mass, and
their ratios differ.
Mass splits are checked to conserve at four awkward target sizes and four growth
fractions, where every share floors and the last channel absorbs the remainder.
An absent capability is checked to yield a zero budget rather than a default, and
unknown rule versions and out-of-range growth fail closed.
The genome-to-ledger path is exercised end to end: a derived `ConsumerConfig`
drives a real `FeedingCommunity` tick, the ledger totals are identical before and
after, and a bite one femtogram past the derived cap is refused by the existing
M2.3 contract. The opening inventory is checked to be exactly the caller's.
Review of `be71d4f` found one defect (2026-09-23): construction composed the mass
difference rather than differencing the two composed bodies, which discards the
rounding the existing body already carries. At target mass 1001 growing from 10 to
60, it funded C25 N7 P2 other16: the right total, but nitrogen and phosphorus one
short each and other two over, so spending the budget could not produce the body
the same rules derive. Construction is now the per-channel difference between the
composed bodies, which funds C25 N8 P3 other14 for that case.
Because the remainder rule is not monotone per channel, that difference can be
negative on a channel, so construction became a transition with an intake and a
release rather than a single requirement: a body growing from 13 to 14 femtograms
acquires one carbon and one nitrogen while freeing one from other. The release is
a rounding artifact bounded by a few femtograms, not excretion.
A follow-up review found the affordability half of the same problem: a transition
could require more of an element than was available, because the check compared
net mass. A step growing a body from 13 to 14 femtograms nets one femtogram but
acquires a carbon and a nitrogen while freeing one of other, and released other
pays for neither. The environment now states availability per ledger channel
rather than as one pool, since a pool cannot say that carbon is plentiful while
nitrogen is not. `plan_development` keeps a total-mass bound, sound because net
never exceeds gross but not an authority, and `metabolic_budget` narrows it to the
largest step whose own intake fits channel by channel and reports the growth that
step funds. That revises ADR-0031 and is recorded there.
The affordability tests place a single femtogram in each channel in turn and
require that no step needing a different element is funded, that nothing is
acquired beyond what is present, and that the mass-only plan would have admitted
the step; a carbon and a nitrogen together then fund it exactly. Across six target
sizes by seven growth fractions by four stock levels, every reported intake is
checked to fit the availability channel by channel and never to exceed the step
the plan proposed. These tests were confirmed to fail with the per-channel check
removed.
A third review found the availability bound itself could wrap: summing four
unrestricted u64 channels with 2^62 in each reaches exactly 2^64, so the total
read as zero and both planning and budgeting refused growth that a billion
femtograms per channel funded. The sum now saturates, which is also the right
answer since no step can need more than u64. Regressions in both suites use that
exact 2^62 case and a single saturated channel, and require the same step a
billion per channel funds; both were confirmed to fail without saturation. The
equivalent helper in `metabolism.cpp` is applied only to inventories derived from
bounded phenotype masses, never to caller input, so it is left unsaturated rather
than given a branch no test could reach.
The invariant `held + intake - release == reached` is now asserted per channel for
the reviewer's exact case, for the releasing case, across six target sizes by
seven growth fractions by four material budgets, and across a forty-step walk from
nothing to full growth where the accumulated intake and release are checked
against the whole history. The net of every transition is also asserted to equal
what `plan_development` charges material for. These tests were confirmed to fail
against the previous implementation.
Coverage limitation: as in M3.2, the overflow guards are unreachable through the
public API at schema 1 ranges and are recorded rather than covered by a test that
could not fail. The small integer helpers are duplicated between `development.cpp`
and `metabolism.cpp` rather than shared through an internal header.
No organism spends any budget: nothing consumes maintenance, movement or
construction, and no starvation, death or repair rule exists. The constants are
chosen for legible magnitudes, not balance. Golden `d61ad5bf0c928c76`, ecosystem
fingerprint `2cd4a15f9fbe9949` and the frozen genome fixture are unchanged. No
deployment manifest or image pin changed. Linux validation is pending its first
pushed run; no Gitea result is claimed for any increment since `b01a824`.
## M3.4 seeded mutation candidates (2026-09-23)
Native Release, analysis and MSVC ASan each passed 20/20 CTests; formatting
passed. Commands: `tools/windows-check.ps1 -Preset <release|analysis|asan>
-Action check` and debug `-Action format`. `mutation-contracts` is the one new
CTest case, in the existing `evo::biology` library. No existing suite changed.
The eligible set for the founder fixture is pinned by hand rather than observed:
nine operators, derived from the schema by the rules that make them eligible.
Energy conversion and membrane can adjust either way, movement, ingestion,
reproduction and thermal tolerance can be acquired, target size can adjust up but
not down because its 9,999,999 step would fall below the minimum from 8,000,000,
and no required capability can be shed. Over 2,000 seeds the offers are exactly
that set and nothing else.
Uniformity is checked to within a fifth of each operator's exact share, about six
standard deviations for 8,000 offers across nine operators. That test was
confirmed to fail when the draw was replaced by taking the first eligible
operators, which is the shape a biased offering would take. The rejection
sampling that removes modulo bias is correct by construction rather than
test-verified: over nine bins with 64-bit draws the bias it removes is about one
part in 10^18 and no test at this scale could observe it. That is recorded rather
than claimed as covered.
Stability is checked both ways: the same context always returns the same
candidates, while a new event index, seed, stream or growth state returns a
different draw. Selection is checked to refuse an operator that was never offered,
a moved event index, a changed growth state, a tampered offer and a different
parent genome, each with its own error. Boundary operators are checked to be
absent rather than clamped at both ends of a range, and a capability another one
declares as a prerequisite is checked to be unsheddable while the dependent
capability itself remains sheddable. Every offered candidate is selected in turn
and its successor checked to validate, to record parent 7, and to survive a
decode; the reported maintenance and mass are checked to equal what
`metabolic_budget` charges that successor.
Review of `5497f4d` found one defect (2026-09-23): selection established a
stale context by comparing regenerated candidate lists, which cannot establish it.
Four of nine operators is only 126 distinct lists, so different contexts draw the
same list routinely, and an offer from event 1 was accepted under event 239. An
offer now records the parent's identity and digest and the context itself, and
selection compares those directly before anything else. Because they then
determine the draw, a list that still differs means the offer was altered, which
is now its own error rather than a context having moved.
The regression searches for a context whose draw collides with the offer's,
asserts that such a collision exists rather than assuming it, and requires that
selection still rejects it. It was confirmed to fail against the previous
implementation. Rewriting the offer's recorded context to name the colliding one
is checked to be accepted, and should be: those candidates genuinely are what that
context draws, so the result is a legitimate offer for it rather than a forgery.
Editing a candidate is checked to be caught as an altered offer.
No organism mutates and no fixture calls any of this. Candidate count, context
window, mutation cost as a game rule, inherited versus acquired expression, NPC
mutation and recombination remain open and are not settled here. Golden
`d61ad5bf0c928c76`, ecosystem fingerprint `2cd4a15f9fbe9949` and the frozen genome
fixture are unchanged. No deployment manifest or image pin changed. Linux
validation is pending; no Gitea result is claimed since `b01a824`.
## M3.5 lineage retention (2026-09-23)
Native Release, analysis and MSVC ASan each passed 21/21 CTests; formatting
passed. Commands: `tools/windows-check.ps1 -Preset <release|asan|analysis>
-Action check` and debug `-Action format`. `lineage-contracts` is the one new
CTest case, in the existing `evo::biology` library. No existing suite changed.
Retention is asserted as a property of the type, not only of its behaviour: four
concepts check that the archive offers no `erase`, `remove`, `clear` or `prune`,
with a std::map positive control so the checks are known to be live rather than
vacuously true. An earlier attempt used bare requires-expressions, which MSVC
diagnoses eagerly for a non-dependent type, and a std::vector control, which was
wrong because `vector::erase` takes an iterator rather than a key; the concept
correctly reported false and the control was corrected rather than the assertion
weakened.
A branching fixture of five genomes over three generations is used throughout.
Tests cover a genome outliving both organisms that carried it with its record,
ancestry, children and organism terminal ticks intact; ancestry and descendants
resolving through a genome whose only carrier died; common ancestors for
siblings, cousins, a genome against itself and two separate founders; refusal of
a duplicate identity, a child admitted before its parent and a full archive;
death being terminal against repeat, reversal and rebirth; and absence being
reported for every lineage that has no living carrier, including one that gains a
carrier again afterwards. The capacity case admits the full 4,096 genomes as one
chain, checks that the refusal leaves every earlier record intact, and walks the
4,096-deep ancestry.
The implementation was changed twice under clang-tidy rather than suppressed.
`std::map` members made the archive's move potentially throwing, because MSVC's
map move constructor allocates a sentinel node; sorted vectors replaced them,
which also matches the rest of the codebase. A wrapper struct holding a `Genome`
by value then had the same problem, because ADR-0030 made `Genome` copy-only, so
genomes are now stored directly and identified by their own `id()`. The redundant
per-genome carrier tally was dropped in the same change: carriers are counted from
the organism records, so there is no second tally that could drift.
Review of `8020347` found one defect (2026-09-23): the retention claim covered
methods named like removal and nothing else, so `archive = LineageArchive{}`
discarded every record, and assigning a copy taken before a death restored that
organism's living state. Both forms of assignment are now deleted, which also
makes the archive unswappable, and the copy constructor is declared so the
implicit moves are suppressed and a move decays to a copy. Copying stays available
because a clone removes nothing and GENOME_MODEL expects cloned-world rehearsals.
The interface tests now cover assignment, replacement from a temporary and
swapping as well as the named methods, each with a std::vector positive control,
and check at runtime that a copy taken before a death does not disturb the
original's terminal record and that a moved-from archive still holds everything.
All four static assertions were confirmed to fail with the deleted assignments
removed.
Linux CI on `29ce255` reported one build blocker (2026-09-23), the first Linux
result recorded since `b01a824`. GCC Release passed 21/21, then the Clang 19
analysis build failed: a bare `{1, 2, 3, 4, 5}` deduces `initializer_list<int>`
and the sign-changing conversion to each `std::uint64_t` element trips
`-Wconversion` under `-Werror`. The literals are now unsigned.
MSVC `/W4 /WX` does not diagnose this, and GCC's `-Wconversion` does not warn when
a constant value is known to fit, so neither the local checks nor the GCC job
could have caught it; only Clang did. Attempting to reproduce Clang's diagnostic
locally through clang-tidy failed, because the warning flags are not honoured in
its cl driver mode and no clang driver is installed on this host. The whole tree
was instead audited by hand for the pattern: every other range-for over a braced
literal list either deduces through `auto` or already carries suffixes matching
its declared type, and this was the only instance. That audit is a one-off, not a
check that will catch the next one; the standing gap is that no Clang or GCC
compile happens on this host before a push.
Nothing is serialized, so no history survives a process, and no fixture calls any
of this. Terminal extinction is not claimed anywhere, durable retention and
migration are M7 and M8, and NPC mutation remains undecided. Golden
`d61ad5bf0c928c76`, ecosystem fingerprint `2cd4a15f9fbe9949` and the frozen genome
fixture are unchanged. No deployment manifest or image pin changed. Linux
validation is pending; no Gitea result is claimed since `b01a824`.
## M3.6 genome ecology ensembles
Executed on this Windows host (2026-09-23): Release configure, build and `ctest`
33/33; the analysis preset (clang-tidy 19, warnings as errors) built clean; the
ASan preset configured, built and ran 33/33; `clang-format` 19.1.5 reported no
violations after the new files were formatted. All four checks ran and passed on
the committed tree. No Clang or GCC compile happens on this host, so the standing
gap recorded for `29ce255` is unchanged and Linux CI remains the first Clang and
GCC result for this increment.
`evo-ensemble-tests` covers eight properties: that a derived strategy carries
exactly ADR-0032's budget quantities and invents no margin, that divergence is
reproducible and refuses its bounds rather than truncating, that conservation
holds at every tick of every cell, that a run is reproducible and a different seed
is not, that the archive ends holding exactly the founders and every organism that
lived with no living carriers, that the two seed halves agree within the stated
band, that each gradient axis alone changes the outcome, and that a thermally
tolerant strategy keeps more scheduled life at the extremes. `evo-sim --ensemble`
is pinned by fingerprint `c2b96c86101be628`, confirmed stable across three runs
and distinct from a different seed's `8f88e72cf2715eac`.
Three defects were found and fixed during implementation, all by measurement
rather than by review. `EnsembleReport` and `CellReport` were default-initialized
rather than value-initialized: both are aggregates whose vector members give them
a non-trivial default constructor, which leaves the scalar counters indeterminate,
and the episode total read as garbage. Both are now `{}`.
The other two are findings about the model, and are recorded in ADR-0035 rather
than tuned away. The header claimed the resource level entered both the derived
strategies and the episode's nutrient pool. It does not enter the derivation at
all: founders are derived fully grown, so their construction transition is empty
and ADR-0031's per-channel `available` cannot reach intake, assimilation or
reproduction investment. The comment was corrected rather than the code, because
the code was right and the claim was false.
The resource axis was then measured to be decorative at the original constants and
non-monotonic once it was not. A ten-point sweep at the original fixture returned
41 to 45 organisms across three orders of magnitude of nutrients with identical
tick totals and eight extinctions everywhere: every birth came from the opening
endowment, because a body of 353,600,000 fg needed 73,776,000 fg of carbon to
reproduce while feeding supplied 8,000 fg per tick before assimilation, or roughly
23,000 ticks against a 60-tick lifespan. Rescaling so a birth can be earned
required pinning the investment fraction to the genome schema's floor of one per
thousand, because intake scales with surface and investment with body mass, so
the ticks needed to fund a birth fall only as the cube root of mass. That a
reproduction is affordable only at the schema floor is a finding about the
ADR-0031/0032 cost model and is now the second priority in STATUS.md.
After rescaling, a sweep of 200 to 200,000 femtograms gave 45, 45, 45, 51, 61, 81,
44, 44, 42 and 42 organisms: a peak near 8,000 and a fall away on both sides. The
mechanism is ADR-0026's feeding rule that a target may be bitten by only one
consumer per tick, which caps community income at one producer bite however large
the pool; scarcity empties the producer, consumers bite each other, and that
concentrates mass into individuals that reach the threshold, while abundance
spreads the same income too thin to fund a birth. The gradient test therefore
asserts only that an axis moves the outcome, never a direction. An earlier draft
of that test asserted `rich_organisms > lean_organisms`; that assertion held at
the two chosen cells but implied a monotonicity the sweep disproves, so it was
replaced rather than satisfied by picking two cells that flattered it.
Coverage limits. The grid is four cells and the tests use eight seeds per cell, so
every number describes this fixture and nothing wider. The split-half band
(`high * 2 <= low * 3 + 2`) is a chosen tolerance, not a derived one. No organism
mutates during an episode, so nothing here exercises in-episode divergence and NPC
mutation stays open. The thermal penalty lives in the harness and no contract
knows about it, so no contract test covers it. Nothing is serialized.
The eleven new CLI tests are new-behaviour tests for a new mode, not regressions
of a fixed defect; only the M2.6 bidirectional mode check they extend was ever a
regression. They were confirmed to fail the run, which is what `WILL_FAIL`
asserts, but each one was not separately confirmed to fail through the branch it
names, so a rejection arriving from a different branch would still pass. Golden `d61ad5bf0c928c76`, ecosystem fingerprint `2cd4a15f9fbe9949` and the
frozen genome fixture are unchanged, and ADR-0029's episode driver was not
touched. No deployment manifest or image pin changed. Linux validation is pending;
no Gitea result is claimed since `b01a824`.
Review of `89ad95b` found three defects (2026-09-24), each now fixed with a
regression confirmed to fail against the unfixed code.
The harness recorded a death for every organism still alive when an episode
stopped. A one-tick run therefore reported zero extinctions while writing terminal
death records for founders that had not died, and every such organism's watched
time was added to the same total as a completed lifetime. Ending observation is
not death, and ADR-0034 makes a recorded death terminal, so there was nothing for
the archive to be terminal about. Survivors now stay living, are counted in
`survivors`, and their watched time accumulates in `observed_ticks`, which is a
lower bound on a lifetime the run never saw end. `deaths` and `lived_ticks` cover
only organisms observed to die, and the two are never added together. The run and
archive test now checks that deaths plus survivors equals every organism recorded
and that a lineage is absent exactly when all its carriers were seen to die.
A draw that produced a genome without the reproduction capability derived a zero
reproduction investment, which ADR-0029's driver refused outright, so
`--ensemble --seed 10` exited with invalid_arguments; seeds 12 and 15 were
confirmed to fail the same way and seed 11 to succeed. Filtering such variants out
before sampling was rejected as the worse repair: the measured set would then have
been the fertile subset chosen by a guard, and an inability to reproduce is a
result rather than a reason for exclusion. The driver now accepts the strategy and
proposes no birth for it, since an empty investment would otherwise clear both
reproduction bounds trivially and request a birth transferring nothing. ADR-0029
is amended rather than silently contradicted, and ADR-0035's claim that the driver
does not change was corrected to say what did.
Only configurations the driver previously rejected outright behave differently, so
the M2.6 pin was expected to hold and was checked rather than assumed:
`--ecosystem --seed 0 --episodes 32` still reports `2cd4a15f9fbe9949`, and the
ecosystem, feeding and producer suites pass unchanged. The ecosystem test that
asserted a zero investment is rejected now asserts the opposite fact, that such an
episode runs and reports no births.
A half's thermal shortfall was assigned rather than accumulated, so a field
documented as a total across the half's episodes carried one episode's reading.
Two episodes at 15,000 millikelvin reported 15,000 rather than 30,000. It now
accumulates per episode. The tolerance test's expectations were per-episode
readings and are now stated against the episode count, with an assertion that the
count exceeds one so a sum and a single reading cannot look alike.
`cli-ensemble` was repinned from `c2b96c86101be628` to `5ec59d390909b526`,
confirmed stable across three runs and distinct from seed 7's `090b83efb85ae477`.
This is a change in reported facts, not a test accommodation: survivors are no
longer counted as deaths, the outcome record gained `deaths`, `survivors` and
`observed_ticks`, and the shortfall is now a sum, so the previous value describes
a report this build does not produce. The rationale is recorded beside the pin.
Each fix was confirmed to be load-bearing by reverting it alone and rebuilding.
Restoring the survivor-as-death behaviour failed the archive test at the survivor
count and, when that test was run second, the dedicated regression at `deaths ==
0`. Restoring the zero-investment rejection failed the sterile-strategy regression
at `EcosystemEpisode::create`. Restoring the assignment of thermal shortfall
failed the summation regression. One of my own expectations was wrong rather than
the code: a one-tick episode's survivors include organisms born during that tick,
watched for none of it, so observed time equals one tick per founder rather than
one per survivor. The test now says that.
Coverage limits unchanged from the increment above, with two additions. Censored
lifetimes mean a run's mean lifetime is not computable from `lived_ticks` alone
and no test asserts one. Across a run the archive now ends holding living
carriers, which is accurate about what was observed and is not a claim that
anything persists; nothing is serialized.
## M4.1 cohort state and ownership
Executed on this Windows host (2026-09-24): Release configure, build and `ctest`
34/34; the analysis preset (clang-tidy 19, warnings as errors) built clean; the
ASan preset configured, built and ran 34/34; `clang-format` 19.1.5 reported no
violations. All four checks ran and passed on the committed tree. No Clang or GCC
compile happens on this host, so Linux CI remains the first result for the new
`evo::ecology` library.
`evo-cohort-tests` covers six properties: that a cohort keeps exactly the fields
it was given, that totals equal the summed classes, that two cohorts with
identical totals but different structure are distinguishable, that every
construction invariant is rejected by its own named error, that one habitat and
genome holds one record, and that only the recorded owner under a current epoch
may replace it. A sixth checks that a cohort names a genome the archive resolves
ancestry for without storing that ancestry.
Every guard was confirmed load-bearing rather than assumed. Each of the nine was
removed in turn and the suite rebuilt and rerun: the duplicate, stale epoch, wrong
owner, unowned reserve, residual bound, class ordering, empty class and class
range guards each produced a failing run, and removing the overflow guard fails to
compile, because the running totals then have no reader and the Release build
treats that warning as an error. An earlier attempt at this sweep reported all
nine as not load-bearing; that result was false and was discarded. The harness
wrote to `/tmp`, which Git Bash and PowerShell resolve to different directories,
so the mutation script was never found and the source was never modified. The
sweep was rerun against an absolute path.
Two defects in my own test code were found and fixed before the suite was trusted.
The single-argument `CHECK` macro split on the comma inside a braced initializer
such as `Owner{11, 2}`, which is a compile error rather than a silent one, and the
macro is now variadic. More seriously, the rejection tests originally asserted on
`.error()` alone, which is undefined when the call unexpectedly succeeded and would
have let a removed guard pass quietly; they now check the failure first through a
`refused` helper. A regex rewrite applied to make that change spanned lines and
corrupted several positive assertions, so the test file was rewritten rather than
patched.
Coverage limits. Nothing here is serialized, so `schemas/population-v1.md` pins
field meaning and invariants and there is no encoder, no round trip and no
migration to test. The approximation bound is carried and never computed or
checked, so no test asserts that aggregation is faithful; that is M4.5.
Conservation between a cohort and the rest of the world is not asserted, because
this type holds an inventory without participating in any transfer. Ownership is
tested as a single-process invariant only: epochs make a stale writer fail locally
and nothing here exercises fencing between workers, which is M7 and M10. Negative
reserves are not representable at all, since every quantity is unsigned, so what
the suite rejects instead is a declared state that would require one, and two
static assertions record that the underlying types are unsigned. `max_cohorts` is
enforced but not exercised at its limit, since installing 4,096 records would add
nothing to what the duplicate and lookup tests already establish.
No existing contract changed. The proof golden `d61ad5bf0c928c76`, the ecosystem
fingerprint `2cd4a15f9fbe9949`, the ensemble fingerprint `5ec59d390909b526` and
the frozen genome fixture are all unchanged, and the other 33 tests pass
untouched. No deployment manifest or image pin changed. Linux validation is
pending; no Gitea result is claimed since `b01a824`.
Review of `1c8b292` found two defects (2026-09-24), both fixed with regressions
confirmed to fail against the unfixed code.
Replacement checked the ownership epoch and nothing that ordered two writes made
inside one. A record installed at tick 1,000 holding 50 picojoules could be
replaced by its own earlier tick-900 record holding 100, restoring inventory the
installed record had already spent, because within an epoch the two were
interchangeable. A cohort now carries a revision that strictly increases within an
epoch, and a repeated revision is refused as well as an older one, since a write
has to be strictly later rather than merely not earlier. A newer epoch may restart
the series, because re-acquiring ownership begins a new one. The interval boundary
is checked independently of both, so neither advancing a revision nor re-acquiring
ownership is a way to rewind simulated time. A later revision at the same boundary
stays legitimate: that is a correction, not a rewind.
The implicit move emptied a cohort's age classes while keeping its habitat, genome
and owner, so a moved-from record was the invalid cohort `create` refuses to build
and the registry accepted through a move what it rejects through construction. The
type is now copy-only, which suppresses the implicit moves so a move decays to a
copy. ADR-0030 made the genome copy-only for exactly this reason in M3.1; the
precedent existed and was not applied when this type was written.
Both fixes were confirmed load-bearing by reverting each alone. Removing the
revision check fails the new ordering regression, as does removing the backward
boundary check. Removing the copy declarations fails to compile in Release, and
not because of the static assertion: with that assertion also removed, MSVC proves
the code after `source.classes().size() == 3` unreachable, having established that
a moved-from vector is empty and the check therefore always throws. That is a
compile-time proof of the assertion rather than an execution of it, so the runtime
path was checked separately under the ASan preset, where the same mutation exits 1.
Coverage limits are unchanged from the increment above, with one addition. The
revision is a caller-supplied counter, not one the registry issues, so nothing
prevents a caller from choosing revisions that are unrelated to its own write
order; what is enforced is only that the registry refuses a record that is not
strictly later than the one it holds. Issuing revisions, and fencing between
workers, remain M7 and M10.
No existing contract changed. The proof golden `d61ad5bf0c928c76`, the ecosystem
fingerprint `2cd4a15f9fbe9949`, the ensemble fingerprint `5ec59d390909b526` and
the frozen genome fixture are unchanged. Linux validation is pending; no Gitea
result is claimed since `b01a824`.
## M4.2 collapse to cohort
Executed on this Windows host (2026-09-24): Release configure, build and `ctest`
35/35; the analysis preset (clang-tidy 19, warnings as errors) built clean; the
ASan preset configured, built and ran 35/35; `clang-format` 19.1.5 reported no
violations. All four checks ran and passed on the committed tree. No Clang or GCC
compile happens on this host, so Linux CI remains the first such result.
`evo-collapse-tests` covers seven properties: that a plan totals its members
exactly and moves each member's whole reserve, that ages and the per-class reserve
distribution survive the move without redistribution, that every ineligible state
and every malformed request refuses the whole plan by name, that a committed
collapse leaves the ledger's conserved total unchanged while emptying every member
reservoir and crediting the destination, that every rejection leaves both the
ledger and the registry exactly as they were, that one plan cannot be committed
twice, and that a moved-from plan is still whole.
All ten guards were confirmed load-bearing by removing each in turn and rebuilding:
the ineligible, foreign, duplicate, self-destination, age classification, stale
sample, occupied destination, empty class and commit ordering guards, and the
copy-only declaration. Two needed a second pass. The first age mutation did not
compile, which is not evidence about the test, so it was rewritten to make the
class predicate always match, which places every member in the first class and is
detected at runtime. Removing the copy declarations fails to compile in Release
for the reason recorded for the cohort; with the static assertion also removed the
same mutation was rerun under ASan, where it exits 1, so the runtime assertions
are load-bearing independently of the compile-time proof.
The commit ordering mutation is worth naming: moving the ledger transaction ahead
of the destination check is detected, because a rejected conversion then leaves
biomass in a reservoir nothing owns while the members are already empty. That is
the dual-representation failure this milestone is named after, and the ordering is
what prevents it rather than a comment saying it does not happen.
One expectation of my own was wrong rather than the code. A replay of a committed
plan was expected to report `destination_occupied`, but the staleness check ran
first and reported `stale_plan`, which is equally true. Both are pure checks made
before anything moves, so safety was unaffected and only the diagnosis differed;
the destination check now runs first, because "this conversion already happened"
says more about a replay than "the world moved". The ordering rationale is
recorded beside it.
`accounting::ReservoirId` and `accounting::Transfer` gained the defaulted
comparisons that every other value type in the project already carries, so a
plan's legs can be compared. This is additive, no behaviour depends on it, and
ADR-0037's status was corrected to say so rather than claim no M1 change.
Coverage limits. Completeness of the offered member set is not checkable here and
is not checked: nothing establishes that the caller offered every member it had.
Emptying a reservoir is not withdrawing the right to use it, since the ledger has
no removal, so a later transfer could credit an emptied reservoir and no test
asserts otherwise. The before-and-after property is demonstrated across rejections
inside one process; there is no persistence, so a crash between the ledger
transaction and the registry install is outside what an in-memory registry can
show, and durable conversion sequences, escrow and replay stay M7 and M8. Plans
are bounded to 256 members by the ledger's transaction limit and nothing
coordinates several conversions. The ledger-failure and destination-unavailable
branches are reachable only past bounds the tests do not build, so they are
refused-by-construction rather than exercised.
No existing behaviour changed. The proof golden `d61ad5bf0c928c76`, the ecosystem
fingerprint `2cd4a15f9fbe9949`, the ensemble fingerprint `5ec59d390909b526` and
the frozen genome fixture are unchanged, and the other 34 tests pass untouched.
No deployment manifest or image pin changed. Linux validation is pending; no Gitea
result is claimed since `b01a824`.
Review of `1a59b63` found one defect (2026-09-24), now fixed with a regression
confirmed to fail against the unfixed code.
Commit proved the destination reservoir existed and that the registry slot was
free, but never that the reservoir was empty. A conversion into an occupied
reservoir therefore succeeded and mixed inventory it had never sampled into the
cohort's backing store, leaving the cohort's recorded reserve understating what
backed it and the surplus owned by nobody. The reported case was reproduced
against the unfixed code rather than assumed: gathering into reservoir 5, which
the fixture already seeds with seven picojoules, commits successfully and leaves
the ledger holding 67 against a cohort recording 60.
A destination holding anything is now refused. The success test additionally
asserts that the destination balance equals the installed cohort's reserve, so a
cohort's reserve is the whole of its reservoir rather than a part of it, and the
new regression asserts that the refusal leaves the ledger balances, the ledger
action counter and the registry all exactly as they were, then that emptying the
reservoir first makes the same plan committable.
The check runs after the staleness loop rather than beside the other destination
checks. Both run before anything moves, so this is a reporting choice and not a
safety one: replaying a committed plan into a fresh registry would be refused by
either, and that the members no longer hold what was sampled is the more useful
answer than that the reservoir is occupied. The ordering rationale is recorded
beside the check.
The guard was confirmed load-bearing by removing it alone, which fails the new
regression. An earlier attempt to add that mutation reported the guard as not
load-bearing; the result was void and discarded, because a Bash heredoc had
mangled the escaped newlines in the mutation script and the script failed to
parse, so nothing was ever mutated. It is the second time this environment quirk
has produced a false negative in a mutation sweep, and both times the printed
output rather than the exit status is what revealed it.
## M4.3 interval accounting
Executed on this Windows host (2026-09-24): Release configure, build and `ctest`
36/36; the analysis preset (clang-tidy 19, warnings as errors) built clean; the
ASan preset configured, built and ran 36/36; `clang-format` 19.1.5 reported no
violations. All four checks ran and passed on the committed tree. No Clang or GCC
compile happens on this host, so Linux CI remains the first such result.
`evo-population-tests` covers ten properties, and `evo-cohort-tests` gained two
for the amendments. An interval with no flux advances the clock and submits no
ledger transaction. Births change the population and leave the reserve untouched.
Death, predation, immigration and emigration each move biomass once to the named
counterparty with the conserved total unchanged. A loss exceeding the population
or any one inventory channel is refused rather than clamped, including at the
boundary of a single millionth of a member, while taking exactly everything is
vacancy rather than refusal. A population below one member stays a population and
its record stays installed. A declared result that does not match the fluxes is
refused in both directions, for counts and for biomass. Empty and backward
intervals, a foreign owner and a repeated revision are refused. A commit against a
record that has moved on, or a reservoir that no longer holds what was counted,
changes nothing in either the ledger or the registry, and one plan cannot be
committed twice.
The seeded ensemble runs twenty-four seeds of forty proposed intervals each,
deliberately proposing losses unbounded relative to the cohort so that some must
be refused. Every accepted interval is checked for the conserved total being
unchanged and for the cohort's reservoir equalling its recorded reserve. The test
asserts that both the accepted and the refused path actually occurred, so a change
that made everything refuse, or everything pass, fails rather than looking clean.
All fourteen guards were confirmed load-bearing by removing each in turn and
rerunning: the two no-clamp guards, the two mismatch guards, the interval, owner,
revision and counterparty guards, the three commit staleness guards, and the three
amendments to the cohort record. The two clamp mutations are the ones worth
naming, because they are the defect rather than a missing check: replacing the
refusals with saturating subtraction, which is what a clamp is, is detected.
Two of my own expectations were wrong rather than the code. The escrow arithmetic
in the migration test was miscalculated: the reservoir receives two hundred matter
and returns one hundred, so from seven it stands at a hundred and seven, not a
hundred and ninety-seven. And the dwindling-population test shed all but one
femtogram of the reserve while declaring a result that held none, which is a real
`biomass_mismatch`; the test now sheds the whole reserve. That second one surfaced
a genuine tension worth recording: a class reduced to a fraction of a member may
hold no reserve, because ADR-0036 gives reserve to members and a fraction is not
one yet, so that biomass has to go somewhere in the same interval. It is recorded
in ADR-0038's costs rather than worked around.
ADR-0036 was amended twice, and both were forced by this milestone rather than
chosen. Threading the reservoir through each request instead of binding it to the
record would have let a caller advance one cohort against another's store, and a
balance comparison cannot distinguish two equal reserves, so the binding is a
correctness fix and not tidiness. Vacancy is required because a decline to exactly
zero was otherwise unrepresentable and the only way through would have been the
clamp this milestone exists to refuse. `Cohort::vacant` is a separate factory
precisely so that an empty class list in a draft stays an error.
Coverage limits. Closure is checked and stability is not: an interval whose
arithmetic balances and whose ecology is nonsense is accepted, and no test asserts
otherwise, because comparing matched ensembles is M4.5. The declared age structure
is not verified to be what ageing would produce, since that needs the within-class
distribution the record discards; nothing here checks that members did not get
younger. Emigration escrows into a reservoir nothing settles with. Ownership
remains a single-process invariant. The overflow and ledger-failure branches are
reachable only past bounds the tests do not build, so they are refused by
construction rather than exercised. Nothing is serialized.
The proof golden `d61ad5bf0c928c76`, the ecosystem fingerprint
`2cd4a15f9fbe9949`, the ensemble fingerprint `5ec59d390909b526` and the frozen
genome fixture are unchanged, and the other 33 tests pass untouched. No deployment
manifest or image pin changed. Linux validation is pending; no Gitea result is
claimed since `b01a824`.