An empty reservoir is not an unowned one

Two defects from review of 380d609.

Materialization treated an empty reservoir as available. A vacant cohort, or one
holding nothing, still owns its store, so a draw could credit it while the owning
record went on recording nothing: the same inventory in two places, reached from
the opposite direction to the destination defect in 1a59b63. Destinations are
checked against the registry's ownership now, not just their balance, and the
registry answers whether a reservoir is used rather than callers inferring it from
what the reservoir contains.

The same hole existed on the collapse side in a worse form, found by asking
whether the reported defect had a sibling. Nothing verified the destination was
unowned, so the shared-reservoir rule added in M4.3 would have refused the install
only after the transfer had run, leaving the members emptied and their biomass in
a reservoir nothing accounted for. That is exactly the stranding the M4.2 ordering
was chosen to prevent, defeated by a check that ran too late. Whichever registry
step follows a transfer is now proved to succeed before it, through may_install
and may_replace queries that install and replace delegate to, so there is one rule
and not a restatement that can drift.

The second finding was a functional gap the oscillation test hid by discarding its
registry each cycle. A full draw leaves a vacant record, which ADR-0036 retains
because absence is not extinction, and collapse then refused that habitat and
genome as occupied: a full materialization was a one-way door out of a habitat and
the lineage could never return. Collapse repopulates a vacant record in place now,
continuing its ownership and revision series rather than restarting them, and a
record still holding a population is refused, because merging two silently is not
a conversion. The test keeps one registry for all twelve cycles and asserts the
revision continues.

Four of the five new guards are load-bearing. The fifth is not and is recorded as
such: the may_replace pre-check in commit_materialize is unreachable given the
checks before it, since the installed record is already proved equal to the plan's
origin and the result differs only by a later revision. It is kept because it is
the right place for the guard and its collapse-side sibling is genuinely
reachable, not because it is exercised.

Release 37/37, analysis, ASan 37/37 and clang-format all ran and passed here.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
This commit is contained in:
dingo
2026-09-24 17:47:37 -04:00
co-authored by Claude Opus 5
parent 380d609f06
commit 14e0c1af5e
13 changed files with 332 additions and 37 deletions
+50
View File
@@ -1192,3 +1192,53 @@ fingerprint `2cd4a15f9fbe9949`, the ensemble fingerprint `5ec59d390909b526` and
the frozen genome fixture are unchanged, and the other 36 tests pass untouched. No
deployment manifest or image pin changed. Linux validation is pending; no Gitea
result is claimed since `b01a824`.
Review of `380d609` found two defects (2026-09-24), both fixed with regressions
confirmed to fail against the unfixed code.
Materialization treated an empty reservoir as an available one. A vacant cohort, or
one simply holding nothing, still owns its store, so a draw could credit that store
while the owning record went on recording nothing: the same inventory in two places,
reached from the opposite direction to the destination defect found in `1a59b63`.
Destinations are now checked against the registry's ownership as well as their
balance, and the registry answers whether a reservoir is used rather than callers
inferring availability from what it contains.
The same hole existed on the collapse side in a worse form, and was found by
checking whether the reported defect had a sibling rather than by being reported.
Nothing verified that the destination reservoir was unowned, so the shared-reservoir
rule added in M4.3 would have refused the install only after the ledger transfer had
already run, leaving the members emptied and their biomass in a reservoir nothing
accounted for. That is exactly the stranding the M4.2 ordering was chosen to prevent,
defeated by a check that ran too late. Whichever registry step will follow a transfer
is now proved to succeed before it, through new `may_install` and `may_replace`
queries that `install` and `replace` themselves delegate to, so there is one rule
rather than a restatement that could drift.
The second finding was a functional gap the oscillation test had been hiding by
discarding its registry every cycle. A full draw leaves a vacant record, which
ADR-0036 retains because absence is not extinction, and collapse then refused that
habitat and genome as occupied. A full materialization was therefore a one-way door
out of a habitat: the lineage that had been there could never return. Collapse now
repopulates a vacant record in place, continuing its ownership and revision series
rather than restarting them, and a record that still holds a population is still
refused, because merging two silently is not a conversion. The oscillation test keeps
one registry for all twelve cycles and asserts the revision continues across each
repopulation, so the round trip is exercised against the world it left behind.
Four of the five new guards were confirmed load-bearing by removing each alone: the
ownership check on both sides, the collapse pre-check ordering, and the repopulation
path. The fifth is not, and is recorded rather than presented as tested. The
`may_replace` pre-check in `commit_materialize` is unreachable given the checks
before it: the installed record has already been proved equal to the one the plan
was built from, and the plan's result carries the same habitat, genome, reservoir and
tick with a strictly later revision, so the replacement cannot be refused. Removing
it changes nothing the tests can see. It is kept because it is the structurally
correct place for the guard and its sibling on the collapse path is genuinely
reachable, not because it is exercised.
An environment note for the record: a Bash heredoc containing apostrophes failed to
parse, so the larger test edits in this increment were written to a scratchpad file
and run from there. That is the third distinct way heredocs have mangled content in
this project, after backslash escapes and `/tmp` resolving differently between
shells.