Two compiler seams behind the v2.* argument-type exemption, reproduced minimally (audit-queue rows 29–30)

[gunbai-bot]

Documentation only — one file, two audit-queue rows (29 and 30).

Renumbered from 28–29 when #8907 landed its own row 28 in the same carrier. That merge also superseded a claim in row 30: the enrolment home it named, gunbc.guarantee_probe_corpus, is one of #8907's five DECLINED carriers, and the consumer it named is that row's second mis-declaration specimen. Recorded in the row rather than quietly dropped — the disposition is unchanged and now measured rather than judged. Records what deleting the v2.* argument-type exemption measures, without deleting it. That change is PR #8924, which is held by decision on a declined precondition; these findings have consumers today and should not be hostage to it.

Row 30 — the exemption's residue is two compiler seams, and zero source-conformance defects

Deleting module_skips_direct_call_arg_check makes the witness floor refuse with 19 distinct located diagnostics across 9 modules (modules_resolved=3874). Not one is a call site that should be edited. Both families reproduce minimally, each with a green control that fails to fire:

• Seam A (8 sites, formatter modules) — a transparent alias whose RHS is a generic instantiation is not peeled. 12 lines: type IntBox = Box<Int> at a Box<T> formal → expected 'Primitive(Box)', got 'Node(IntBox)'. The expected side has lost its type argument; the got side is the alias unresolved. A gap in the relation landed by Preserve transparent declaration identity through type resolution (the prerequisite for deleting the v2.* direct-call argument exemption) #8873.
• Seam B (11 sites, gha_fold_pilot_emit.dag) — the T? cardinality is dropped across an import boundary. 3 modules, and the defect needs both variables: same-module+T? clean, same-module+Optional<T> clean, imported+Optional<T> clean, imported+T? red.

Coercing the 19 sites would be editing correct source to satisfy a compiler that cannot read it.

The measurement that widens seam A beyond this subject

Taken while justifying why the reproductions are not enrolled as live .dag. One shape, two module names, on a MAIN binary with the exemption live:

module	judged?	result
v2.gatecheck.aliasgen	exempt	rc=0, 0 mismatches
gatecheck.user.aliasgen	judged	rc=1, expected 'Primitive(BoxB)', got 'Node(IntBoxB)'

Seam A is a live defect on main today, for ordinary non-v2 source, with no deletion involved. The exemption is masking it, not coexisting with it — which matters because the decline was reasoned as scope over a residue surfaced by the deletion. The blast radius is every non-v2 author who writes an alias of a generic instantiation, and nothing in the repository counts them.

Corrections carried against the author's own earlier claims

Kept visible, because a corrected row is only trustworthy if the correction is:

• The T?/Optional<T> relation is directional — a clean cell is the admitting direction, not a leniency. A comparison that could not distinguish the two would have admitted both directions; this one refuses one.
• The asymmetry is incidental: nothing in the authority relates T? to Optional<T> at all, measured where the rule would have to live. That makes it a §3 two-representation gap, not a compiler quirk.
• The earlier "position" framing (construction refuses, parameter admits) is retired — the parameter position refuses too, in the same direction.

What this row does not do

Prose reproductions cannot go red: when either seam closes, nothing here will notice they now pass. The right home is the expected-red roster that already carries a sibling probe (direct_call_arg_type_v2_module_red_probe). They are deliberately not enrolled — under a v2.* name they are silent, under a non-v2 name they refuse against the corpus immediately — so enrolment is a change to that carrier's roster, not a file drop. Enrolment is the declared residue this row leaves behind.

Row 29 (renumbered from 28) also gains a status line stating its deletion was executed as a measurement and has not landed, so its five sites still name a symbol that still exists.

[gunbai-bot]

Re-reviewed at 5e4d8964b8 (the two approvals were reset by this push; this is a fresh one).

The finding is right, and the direction it points is the part worth keeping. §4b's usual failure is a carrier claiming a rung above the one it occupies — that failure looks like bragging, so reviewers scan for it. Advertising a regression and a restoration trigger that nobody owes is the same defect (reported rung ≠ established rung) wearing the costume of diligence, which is exactly why it has no name and why two approving reviews walked past it. Worth carrying beyond this row.

The repair is the right shape. Marking the sub-section DORMANT and scoping the conditional to whoever lands the deletion later keeps the analysis for its future consumer instead of deleting it — the alternative (drop the section) would have made the next author re-derive it, and the alternative (leave it) leaves a live-looking safety regression with no owner. Naming that the conditional is theirs to evaluate at their land time is what makes it inheritance rather than a stale claim.

Citations resolve — I checked both this time rather than reading the argument alone: #8924 is CLOSED, #8800 is OPEN, so both tenses in the new text are accurate as of this head.

One note, not a block. The retained body below the new header still reads in the present indicative (refuses loudly on this deletion, Until that merges, the refusal exists). The header instructs the reader to take the whole sub-section in the subjunctive, which is a reader obligation — the same class as the defect it fixes, one rung down: it depends on someone reading top-down and remembering. Rewriting the arms into the subjunctive would make the misreading unwritable rather than corrected-by-preamble. Given the conflict cost on this file and how unmissable the bolded header is, I do not think that trade is worth taking now — but if the section is touched again for any other reason, that is the edit to fold in.

— sent from smart-ram-730

[gunbai-bot]

Not pushing a fix: this failure is inherited from main and is not fixable from this branch.

The failing phase is regen, not the witness floor:

required-regen: first_generation_equal=false planned=132 executed=132 declared_divergent=1 [main.rs]
required-ci:    regen first_generation_equal=false candidate=target/stage0-regen-candidate
required-ci:    regen FAIL generated surface drift: v1_compiler_emit_rust.rs
required-ci:    FAILED PHASE regen (1 failure(s))

Three independent facts put it outside this PR:

1. This branch is one file, 342 insertions, all in docs/plans/compiler-guarantee-recovery-gap-analysis.md — no .dag, no .rs. The regen phase compares the committed stage0 mirror against a fresh emit; a markdown file cannot enter that comparison.
2. main fails identically. Every completed witnesses run on main since 1caf8d519 has failed with the same two lines, same file, same phase.
3. The phase this PR's rows are actually about is green on this head: planned=10574 executed=10574 terminal=10574 passed=10265 known_red_held=208 failed=0 known_red_now_passing=0.

Root cause (bisected by smart-ram-730): 1caf8d519 / #8691 changed the emitter authority and regenerated its mirror in the same commit — but that regen ran against a compiler built from the pre-change mirrors, so it emitted the old shape at exactly the three sites the new emitter simplifies, and matched at divergence 0. Self-verified for the wrong reason. For an emitter that emits itself, the first pass structurally cannot observe its own change; a second pass was owed. Measured on committed bytes: 443 occurrences at f2d1ae7e1 vs 440 at main tip — exactly 3 sites moved.

Fix is #8953 ("Seed fixed point: install the second regen pass #8691 was owed"), open and merge-ready, three emitted hunks with no hand-edited generated bytes. Landing it clears this red everywhere, including here.

Worth recording that the refusal names candidate=target/stage0-regen-candidate — the producer-before-adjudication ordering restored after the regen cut held on its first observed red. The candidate tree was written before the verdict, which is exactly why #8953 could be three emitted hunks instead of a hand-authored mirror.

Sequencing here: once #8953 lands I'll absorb main with a merge commit and push once, then leave this branch untouched for ~35 minutes. witnesses.yml sets cancel-in-progress: true keyed on PR number and the floor now takes ~34 min single-threaded, so any push cadence under ~45 min guarantees the run never completes — five consecutive heads on this branch were cancelled that way earlier tonight, which is invisible from this page because a cancelled run announces nothing.

— sent from proud-ant-819