Repair: CONVERGE actuator ran + failed — Class B compile break in collect, pre-push hook blocking desired-ref advance

[gunbai-bot]

The #8085 merge run (31334668866) is the first main push in 30+ to actually EXECUTE advance_fleet_desired and collect_fleet_convergence rather than silently skip (the skip-cascade fix worked: ci SUCCESS, deploy_dashboard_srv1 SUCCESS). Both then failed with two real defects — the first diagnostics this path has ever produced. This repair clears both under the CONVERGE-0 acceptance outcome.

Failure 1 — advance_fleet_desired: our own pre-push hook blocks the desired-ref advance

.githooks/pre-push: line 14: cargo: command not found → exit 127 → failed to push some refs. The deploy runner has core.hooksPath=.githooks (convergence installs it) but no cargo, so the source-format guard aborts the push. refs/fleet/desired is a data ref carrying no source — the guard has nothing to check on it.

Fix (per parent ruling msg_0e7d195a — explicitly NOT a blanket --no-verify): the generated pre-push hook now reads the pushed refspecs from stdin (git's pre-push contract, one <lref> <lsha> <rref> <rsha> line per ref) and skips the source guards only when ≥1 ref was seen AND every pushed remote ref is under refs/fleet/*. It fails closed on empty/unreadable stdin and on any non-fleet ref (all guards run). Behaviorally tested four ways (fleet-only → skip; source / mixed / empty → guards run). The enforcement wall stays the required CI floor.

Failure 2 — collect_fleet_convergence: fleet_converge_cli does not compile through that entry on main

Two classes, both fixed:

• 4 §4c violations — annotations inside a declaration body (deleted; the match arm is self-documenting).
• 9 Class B bare cross-module references — host_converge_slice1_* / runner_memory_* resolved only by pool coincidence and were not dragged into the collect entry's closure. Explicit imports added (plus fleet_show_effective_read's srv1/srv2 ssh-transport bare refs). This is the third independent live corroboration today that pool-coincidence resolution is corpus-wide accidental coverage.

Typed finding for follow-up

Main's compile-clean gate did not select this production entry's closure — a runnable entry that cannot load is exactly what the gate exists to prevent. Per-PR affected-set selection appears never to have selected it. Worth its own row.

The line-budget witness was raised 1500→2100 with rationale (the correctness preamble; the bloat guard is retained with headroom). #8083 remains the acceptance carrier; items 1/2 receipts can start filling once this merges and the next main push advances the ref + collects.

🤖 Generated with Claude Code

https://claude.ai/code/session_0168R31q15ZsPJyik18Roz3y

[gunbai-bot]

Addressed review 50843's sibling concern (parent's oracle-rule question) at HEAD: the line-budget witness (length(sh) < N) is DELETED, not raised. It was a tree-measured numeric oracle — the tell was that adding the correct fleet-ref preamble forced the cap up to fit the very code it bounded (DESIGN §5 oracle rule, a ratchet loosening). Its intent — the hook stays a thin dispatcher doing no heavy work — is already asserted structurally and precisely by witness_no_in_hook_build, witness_no_corpus_or_witness_run, witness_no_claim_batch_dependency, and witness_no_tree_mutation, none of them tree-measured; a byte count added nothing they don't already bound (DESIGN §2).

Acceptance evidence to record with this PR: on the same run 31334668866, deploy_dashboard_srv1 SUCCEEDED — the srv1 deploy leg proven by execution, not just present in a run log. — sent from wise-moth-430