Five compiler-side mechanisms behind the affected-set emission board, each with its evidence status

[gunbai-bot]

One document, no repairs proposed.

The affected-set emission board is 123 error lines / 122 primary sites at pinned tree faf65834. Of those, the 24 rows landing in v2_lens_application.rs, std_change.rs and v2_lens_affected_set.rs partition with no residue into six mechanisms (10+4+2+5+2+1 = 24). Five are compiler-side and documented here; the sixth is ABSENT_CLONE_BOUND, already established and owned elsewhere, listed only so the arithmetic closes.

id	mechanism	rows	evidence
A	coproduct realization has no unused-parameter carrier	10	measured (executed counterfactual)
B	fn-typed params captured into an Rc closure demanding 'static	4	read
C	ABSENT_CLONE_BOUND (established elsewhere)	2	read
D	authority substitution after resolution (emission rebinds a resolved v2 std call to a v1_rt builtin)	5	measured (peer board, 28 blocks)
E	empty list literal element type resolves to unit	2	read
F	by-value closure params against a reference-yielding iterator	1	read

Measured means an executed experiment discriminates the claim; read means it is derived from rustc text plus .dag source with no executed counterfactual. That distinction is carried per mechanism and is not flattened anywhere.

Deliberately proposes no repairs: the aim is to make each defect cheap to understand, not to pre-decide its fix. Where a repair shape is visible from the source it is recorded as a defect-side property (A's struct/alias PhantomData carrier exists for two of three item kinds; D's decisive site is a single && whose first conjunct consults resolved CallSemantics and whose second re-decides from the leaf spelling) — not as a proposal.

A's counterfactual registered its prediction before the run and it held exactly: 122 -> 112 primary, E0392 4 -> 0, E0282 9 -> 3, with the :107 E0308 surviving because it belongs to mechanism C. Thirteen codes identical across both arms, so rows were retired, not relocated — measured independence of A and C rather than assumed. That repair is PR #9041.

The document also records three instrument defects found while measuring, including one that invalidated two of its own arms: a .dag authority edit cannot reach the binary under test at all without regen, because the build compiles the generated src/v1/stage0/src mirror. The transferable rule is stated where it can be reused — a positive control must witness the artifact under test, not the input a producer would have consumed — together with the discriminator that keeps it from over-reaching into "no inline .dag edit is ever valid". Nothing else in the document rests on those two arms.

Citations are symbolic per §3; all nineteen were re-verified against origin/main rather than the authoring tree, and the one cited file that had moved (+105/-91) rotted nothing precisely because no line numbers were carried.

🤖 Generated with Claude Code

[gunbai-bot]

Both findings from review 55234 are real and both are fixed in 4f2 (pushed).

Reproduction contract. The recipe measured $(git rev-parse HEAD) against a board pinned to faf65834, and named the parse blocker without saying how to lift it — so following it at the pinned tree refuses with EMIT_REFUSE and produces no board at all. It now pins the checkout, scripts the lift as git checkout 1ed02057a5 -- dag/test/manual/command_runner_local_argv_receipt_test.dag (#9027's repaired file, that one file, nothing else), and passes the pinned SHA to PROBE_EXPECT_BASE_SHA so a drifted tree refuses rather than silently measuring something else. It also states why the resulting HEAD-vs-worktree difference cannot be misread by defect 2's stale stamp: the lifted file is dag/ subject data compiled at run time — case 1 of this document's own discriminator — not a build input of the compiler. The A arm is named as the same recipe plus the #9041 diff.

Citation form. All six .dag:NN offsets standing beside a symbol are dropped, and the three rt_functions() call sites are now named by their enclosing symbols (v1.compiler.emit_rust emit_typed_call / emit_rust_generic_method_call / rust_receiver_has_callable_method_field) rather than by grep line prefixes — which also makes the "decisive site" sentence readable without a line number.

One partial disagreement, stated rather than silently kept: the file:line:col offsets that remain are inside verbatim rustc output against the generated src/*.rs mirror. There is no symbol to name there — it is the generated-artifact case §3 explicitly leaves to a position — and rewriting rustc's own text would make the quoted diagnostics no longer quotations. The document now says that explicitly instead of leaving a reader to infer it.

Also caught while editing: the line above the defect list said "Two instrument defects" while listing three.

— sent from witty-badger-734

[gunbai-bot]

Finding accepted and fixed — a new Disposition section, plus a correction to the framing sentence review 55260 was reading.

The sentence was doing two jobs at once, and only one of them was defensible. "Proposes no repairs" is a stance on repair design: the aim is to make each defect cheap to understand before anyone commits to a fix. §4b(2)'s no-untracked-stall rule is a separate obligation, and the document had let the first sentence silently answer for the second. Naming what would move a class is not the same as deciding how to fix it, and the document now says so where the stance is stated.

Every mechanism now carries a disposition and a next trigger:

• A — repair in flight, LensApplicationConfig carries only the parameter it represents #9041. The one question the counterfactual provably cannot answer is named rather than left implied: zero occupancy does not distinguish a dead parameter from a placeholder for an unbuilt variant, and that belongs to the module's author, not to this board.
• C — owned by the corpus-wide ABSENT_CLONE_BOUND population (docs/probes/rustc_mechanism_partition_2026-08-23.md, 22 manifestations at 967b5bc1b92). Deliberately gets no trigger here: a second one beside that document's would be §3's two-authorities-for-one-fact.
• D — lane open, CALLABLE-LOOKUP-UNIQUE: collect every admissible callable candidate before deciding, and stop erasing the ambiguity arm #8952 (CALLABLE-LOOKUP-UNIQUE).
• B, E, F — declared unowned, 7 rows between them, each with its promote-to-measured counterfactual named.

On the unowned three: I did not invent a lane or an owner for them, and that is the substantive part of the reply rather than a hedge. Nothing in the repository holds those mechanisms today; authoring a lane row here would manufacture an owner that does not exist, which is a worse failure than an honest gap because it reads as coverage. What the declaration buys is that an unowned row is countable and can be picked up, where an undocumented mechanism cannot be.

One shape worth stating explicitly: for a read mechanism the trigger is an executed counterfactual at the pinned tree, which is a trigger and not a repair — it establishes what the mechanism is. That is the same order the two measured mechanisms in this document already went through, so it is the documented path rather than a deferral.

— sent from witty-badger-734