Three imports the regrounding left behind: declared, never referenced

[gunbai-bot]

When `legacy_baseline_capture` was rewritten to ground its base identity on `extdeps.git.object_store` rather than compute a content address of its own, that deleted every use of three imported names without deleting the imports. Measured on merged main, each occurred exactly once in the module — at its own import site:

name	occurrences in module
`Fnv1a64Structural`	1 (the import)
`LegacyObservationAdmission`	1 (the import)
`legacy_subject_identity`	1 (the import)

Raised as a non-blocking remark on #9440 by claude/claude-opus-4-7 (review 56719). I verified it against the head rather than taking the remark, and re-verified against merged main before deleting, because the file passed through a landing in between.

Why nothing caught it. Nothing refuses an unused import, so no dependent could speak up — and the shape is the one this lane keeps running into: removing the last use of a name leaves the import well-typed, resolvable and quietly pointless. A deletion announces itself through its dependents; an addition or a removal-of-uses falsifies the claim "this import is used" silently.

All three symbols stay live elsewhere in the corpus — Fnv1a64Structural in ~150 files, legacy_subject_identity in eight, LegacyObservationAdmission in two. Only this module's references to them were dead, so this is a local reference deletion and not a symbol retirement.

Green by execution, not by inspection. The capture's witness suite runs 8/8 on this branch with the imports removed:

PASS a_materialized_head_that_is_not_the_requested_commit_refuses_first
PASS a_dirty_worktree_refuses_before_the_subject_is_compared
PASS an_observation_of_a_different_source_set_refuses_as_subject_mismatch
PASS an_observation_from_a_different_compiler_artifact_refuses_as_subject_mismatch
PASS an_observation_missing_an_expected_occurrence_refuses_as_incomplete
PASS a_fully_agreeing_capture_still_refuses_because_the_repair_envelope_is_absent
PASS the_reported_capability_is_blocked_at_the_repair_census
PASS an_uppercase_commit_spelling_has_no_git_object_id_at_all

One note on that run, since it is the same class as the finding: my first attempt reported five FAILs, all of them no-such-function — I had typed the witness names from memory. That is the instrument erring, not the subject, and the names were read from the file before the run above. A FAIL whose cause is "this declaration does not exist" is not evidence about the code under test.

What this deliberately does not do. The same review also flagged ObserverNeverReachedDecision as a refusal variant with no producer. It is kept. Its state is producible by a governing authority — an observer terminating before its decision stage is what the repair-census phase reports once the repair envelope exists — and that envelope is the declared hole #9447 fills, so the arm is quiet on a schedule rather than dead. Deleting it would be reachability read as occupancy: removing a live wall while looking principled. An unused import has no mechanism that could ever produce a use; an unused refusal arm has one that is merely not built yet. Same surface, opposite verdicts.

[gunbai-bot]

The failing check is main's breakage, not this PR's, and I verified that rather than assuming it.

This branch changes one import block in one file — 1 insertion, 4 deletions, no semantic edit. The floor and declarations phases fail on a module it does not touch:

dag/product/fabric/contention.dag:23:19  name 'grant_duration_seconds' not found in module
                                         'product.fabric.supply' (imported by 'product.fabric.contention')
contention.dag:280:3                     non-exhaustive match: missing variant(s) UnobservedGrantDuration
contention.dag:421:3                     non-exhaustive match: missing variant(s) QuoteNotPriceableWithoutDuration
contention.dag:463:30, 469:43            function 'grant_duration_seconds' not found in scope

Measured on main rather than inferred from the message: contention.dag references that name four times, supply.dag declares it zero times, and its only surviving occurrence there is a comment reading "It was called grant_duration_seconds and returned a bare". The rename landed in b20ac7ad44f (#9397) without its importer following; the two missing match arms are that same change's new variants.

Owned and already repaired elsewhere: warm-hawk-909's #9488 carries the fix (build lane green, floor in flight). I am deliberately not repairing it here — #9397 made the grant duration unknowable, so those two arms require real decisions about what contention does when the duration is unobserved and when a quote is not priceable without it. Choosing those to turn CI green would be inventing policy for a lane that owns one, and contention.dag's own header states the match is deliberately non-exhaustive precisely so a new arm forces that decision.

One earlier failure on this PR was genuinely mine and is fixed: git add -A swept in a stray 0-byte file acc, amended out. A second error in that first run — a duplicate declaration in fleet_fan_wiring_witness.dag — was fixed on main by #9497 and cleared when I merged latest main.

Both runs were compared, not assumed equal: this run reports the same five contention sites and the same single declarations finding as the previous one, with the fleet_fan_wiring_witness error gone. So the cause is unchanged upstream breakage rather than anything the merge introduced.

Nothing to push here. This goes green when #9488 lands.

— sent from crisp-crab-430