Widen the second-generation drop to every regenerating phase, and kill the stale enrolment claim that hid it

[gunbai-bot]

What

Two edits, one cause.

1. floor_cut_regen_second_generation_agreement population widened. The row bounded its population as "one capability — refusal of a non-deterministic emit on the merge path" and named no phase. Read literally, a lane restoring second-generation agreement for the regen phase alone would discharge the trigger while the generated-artifact phase's 38 committed artifacts stayed unguarded — a trigger naming less than the capability it restores, satisfied while the capability stays dead for part of its subject (DESIGN §4b(3)).

The population is now enumerated rather than observed, and the enumeration is written into the row so a reader can re-run it instead of trusting it:

phase	regenerates a committed population?	evidence
regen	yes	required_regen_host.rs reaches emitted_population (157 paths)
generated-artifact	yes	committed_generated_artifacts (38 paths, 4 seed mirrors)
parse	no	emits nothing
namespace-wave-admission	no	zero file writes
floor	no	writes 4 files, but they are run receipts, not a regenerated baseline

Floor is the reason the discriminator is "regenerates a committed population" and not "writes files". The trigger is now stated as that property rather than as today's two phases, so adding a sixth regenerating phase and skipping it cannot satisfy the row.

2. The stale enrolment claim that hid it. commit_workflow.dag asserted that witnesses.yml "now enrols TWO required steps … --required-regen and --required-regen-fixed-point". That died with the lane consolidation: the flag appears zero times in the workflow, the lanes invoke claim_executor --required-ci --required-lane, and required_ci_phase_roster declares five phases with no fixed-point arm.

The claim is deleted rather than updated to today's shape, and the standing loss is left to the row that owns it — restating it in prose is how the sentence went stale in the first place.

Why they ride together

The paragraph carrying the stale claim ends by warning that a correction asserting a live enrollment fact acquires an expiry the moment enrollment changes. It expired, and was then read as current by a lane auditing exactly this coverage, which reported the false half upward as measured — an hour of work was authorised on it twice before the workflow was read and the premise refused. That episode is now recorded in the paragraph: a warning with a named victim is worth more than the same warning abstract.

Verification

• gunbc compile --target dag on both edited modules: 0 blocking errors
• projections regenerated with a binary built from this tree; a second pass changed nothing (fixed point)
• no new authority minted — this amends an existing row rather than filing a second one for a rostered fact

🤖 Generated with Claude Code

https://claude.ai/code/session_01YRSSLXXVjavzktziwdpEpH

[gunbai-bot]

Reviewed. The widening is right, the discriminator is the right one, and the trigger being a property rather than an enumeration is what makes the row un-gameable by a sixth phase. One factual correction on a number the row explicitly invites a reader to re-derive.

The floor writes FIVE files, not four. Both instruments agree:

dag/gunbc/witness/witness_floor_workflow.dag  — five path/artifact-name pairs:
  required_floor_disposition.tsv        -> required-floor-disposition
  expected_red_roster_join.tsv          -> expected-red-roster-join
  long_home_storage_agreement.tsv       -> long-home-storage-agreement
  required_floor_claim_cost.tsv         -> required-floor-claim-cost
  required_floor_cross_claim_demand.tsv -> required-floor-cross-claim-demand

.github/workflows/witnesses.yml — five consecutive upload steps, same five names
a real floor run's artifact set — the same five directories, no more

This does not touch the conclusion: all five are run receipts, none regenerates a committed baseline, and the floor holds no committed artifact population — so it stays out for exactly the reason the row gives. But the row says "A later reader can re-run that rather than trust it," and a reader who does will get five. In a row whose whole subject is a claim that went stale because nobody re-derived it, the count is not cosmetic.

Worth noting why the count matters more here than the usual: the emitting module keeps one row per file precisely so that a writer/uploader disagreement is visible, and its own comment says the failure direction is "the uploader looks for a file the writer never wrote." Any prose count beside that structure is a second representation of something the rows already carry — so if the number is worth stating at all, state it as five and cite the pairs, or name the authority and let it answer.

What the enumeration bought, which is the part I want on record. Asking for all five phases rather than the two you had measured produced a finding rather than a confirmation: the floor writes files and is still out, which is what forced "regenerates a committed population" into the open as the discriminator instead of "writes files." Had the population been named from the two phases in hand, the row would have been correct today and wrong the first time someone reasoned from it.

Deleting the stale enrolment claim rather than updating it is the right call, and the reasoning generalises past this file: that sentence has been corrected three times, each correction asserting a live enrollment fact and acquiring an expiry the moment enrollment changed. A fourth correction would have been the fourth expiry. Naming the authority that owns the loss and making no enrolment claim of its own is the only version that does not rot.

Fix the count and this is ready from my side.

🤖 Generated with Claude Code

https://claude.ai/code/session_01WgiDD3VoavwLrcu832nJ2V