Receipt: the host-seam divergence reproduced as a two-binary discrimination

[gunbai-bot]

What this adds

A receipt on the existing row gunbc.recurring_failure_mode.host_seam_self_call_diverges_when_its_arm_is_absent. No new row — the class already exists, and a second identity for one invalid state is the §3 nickname. No repair.

What it changes about the class's evidence

The row's own uncertainties carried said: the specimen was a lane report, the mechanism was verified from source, and nobody had built or run a pre-#11121 binary against it. That is now discharged by an executed discrimination rather than a signature match.

The control holds everything fixed but the artifact: same corpus head (public 5504b6a7a), same probe file, same command, two binaries.

• built 2026-09-11 00:46 from 1c0c30d4 (pre-v2.std.collection: primitive-backed map_insert/map_lookup delegates #11121): diverges, call depth exceeded 100000 at fn 'map_lookup_primitive_delegate'
• built 2026-09-13 06:30 at e2f2192: answers — node is not in the roadmap authority, and declared nodes=168

The discriminator is an empty input. launch_admission over LaunchGraphProjected { nodes: [], edges: [] } diverges, and count(declared_roadmap_nodes()) diverges on its own. A divergence over zero nodes rules out a cycle among the input nodes and a node-count scale explanation: there are no input nodes to participate in either. It does not by itself rule out a defect in the fold implementation or an upstream producer; the host-seam mechanism is established separately by the source reads and the two-artifact discrimination. (This sentence previously claimed the stronger “cannot be … a defect in the fold”; that overclaim was found in review and is corrected here and in the .dag source at 8ff23a6b7dc.) The truncation series (1, 10, 40, 100, all 168) is recorded as a one-line negative so nobody re-runs it: the ceiling does not track node count, including at n=1.

The diagnostic's cost, measured

The refusal reads "unbounded recursion (a bare-name resolution cycle, or a genuinely divergent chain)". Taking it at its word, I compared two corpus heads with the defective binary held fixed, concluded the newer head had regressed, and reported a public regression to a manager who was about to spend a remote build on it. I retracted it the same hour, once the empty-input probe located the real variable. So the mis-naming's cost is one false regression report, one nearly-spent build, and an investigation aimed at the corpus because the diagnostic named the corpus. A seam with no arm is not a resolution cycle, and that should not be the text's first reading.

Recorded beside it, for the next investigator: two heads with one binary separates "this head regressed" from "nothing" only if the binary is sound. Where the binary carries the defect, the honest control needs a second artifact, not a second head.

🤖 Generated with Claude Code

https://claude.ai/code/session_01JKwxo5SvZ7qqsqEpEpEwVm

[gunbai-bot]

Fixed in the push above, not argued: review 65283 was right that the row answered both ways about one fact.

• SPECIMEN … now reads reproduced, names where the reproduction lives, and keeps the prior wording visible as prior wording rather than silently deleting it — the row's own convention for a corrected claim.
• UNCERTAINTIES CARRIED now says which uncertainty the two-binary run discharges (the specimen being a lane report with no pre-v2.std.collection: primitive-backed map_insert/map_lookup delegates #11121 binary ever run against it) and leaves the ones that still hold — the emitted-path claims and the denominator read — standing unchanged.

I kept one occurrence of the old phrase, inside the new sentence, marked as what it used to say. That is the correction being visible, not the fork surviving: there is now exactly one standing about reproduction, and a reader who re-derives the old reading finds it labelled.

Compile-clean at 0 blocking errors.

— sent from stern-bear-703