FIRSTTOUCH-1: transport_script_wall_compile_red -- live_tree_disposition correction, and a terminal admission finding (not a restoration)

[gunbai-bot]

Summary

This started as a restoration attempt and is not one. Reworked after royal-hawk-392 (parent lane) caught that the original diff's green CI run proved nothing — see below.

1. The move out of test.claim.long. stands, on its own reason. 021d17e6a7 quarantined this module for cost (a 35.1x first-touch swing: 1394ms → 48963ms on unchanged rows, traced to the module_path_index shared-fill trap, docs/plans/floor-shared-fill-ledger.md). That trap is already fixed on main (#8470 / 7e4871be86, warms build_module_path_index_from_witness_roots during floor preparation). So the module's long. name was asserting a quarantine reason that is stale now that (2) below is the operative reason — carrying both is the exact drift the long/-vs-module-name split exists to prevent.

2. The real change: live_tree_disposition corrected from SubstrateInputsOnly to ReadsLiveTree. This witness calls compile_dag_rust_emit_check, which reaches the identical build_module_path_index_from_witness_roots live-tree walk its sibling (compile_diagnostic_census_witness_test.dag) already declares ReadsLiveTree for. That sibling's own note had flagged this as "either a latent mis-declaration or a judgment call," pending the falsifier cadence — since deleted (2026-08-15 floor cut) and never adjudicated. SubstrateInputsOnly was the mis-declaration. This correction stands on its own merits regardless of admission consequences.

3. The finding, terminal — this is what "not a restoration" means. v2.workflow.required_floor's admission test (src/v1/stage0/src/cli_run.rs, run_required_floor) is long_home(module) → DeclinedLongModule; else reads_live_tree → DeclinedLiveTree; else Planned — an OR of two independent decline arms, not one quarantine gate. Moving the module out of long. does not admit it once ReadsLiveTree is (correctly) declared: it just moves the three identities from one decline bucket to the other. Confirmed at identity grain via the RequiredFloorDispositionRow receipt (the mechanism that made this decidable at all), comparing this branch's floor run against a main baseline run at the same head:

   • claims (planned): 9686 in both, unchanged
   • declined_long: 539 → 536 (−3)
   • declined_live: 763 → 766 (+3)

   The three identities moved buckets and never became Planned. This witness cannot be admitted to the required floor while it reads the live tree — independent of #8470. The first-touch cost fix and the admission question are two separate facts about this module; fixing the first does not touch the second.

What this leaves open, honestly

The cost question FIRSTTOUCH-1 originally set out to close (is this witness now cheap under #8470's fix) cannot be closed by any floor run while the admission finding holds — the witness is never planned into one. A standalone claim run is not a substitute measurement: it disables the per-floor-process memo #8470 warms, so it would measure a different, uncached quantity, not this module's floor-fold cost. That would be a fourth false reading of the exact defect docs/plans/floor-shared-fill-ledger.md exists to stop people repeating. So the cost question stays open with a named blocker (this witness's own inadmissibility) rather than being left as a checkbox implying a future floor run will close it.

Scale note, unrelated to this witness but worth surfacing where it'll be seen: at the same measurement, declined_live=766 vastly outnumbers declined_long=536 (763 vs 539 pre-move). The multi-wave, multi-session effort behind this whole floor-restoration lane has been aimed entirely at the smaller of the two declined populations.

Test plan

• gunbc compile --source-root dag --source-root src/v2 --dry-run — clean parse/resolve, both before and after the note rewrite.
• Identity-grain disposition comparison against main baseline (above) — the actual falsifying measurement for the original "restoration" claim.
• Swept for stale test.claim.long.transport_script references — none remain.
• Not closeable by this PR or any floor run: the witness's actual floor-fold cost under #8470, blocked on its inadmissibility while ReadsLiveTree.

🤖 Generated with Claude Code