Bind shared floor preparation to exact subjects

[briansrls]

Summary

• model an exact floor-discovery subject with distinct tested commit and Git tree coordinates, ordered root universes, naming authority, tool identity, and typed request mismatches
• model an opaque prepared floor subject over existing ResolveClosureRequest keys and provider_serve outcomes; no MultiEntryIndex or cold-fallback arm crosses the boundary
• make the existing cross-worker discovery snapshot transport bind the complete exact request, verify payload bytes, and refuse dirty/untracked worktrees or damaged/mismatched snapshots without recomputation
• pin the Rust identity mirror to the modeled identity with a shared golden digest

This is the model-first precursor for ci-scoped-worker-shared-substrate. It intentionally does not move run_walk scheduling into .dag because that remains gated by the declared executor-seed deferral, and it does not claim the scoped typecheck world has been removed yet.

Test plan

• both aggregate .dag controls returned true after merging repaired main
• ctrl-build --remote -- cargo test -p v1-compiler --lib cli_run::floor_discovery_snapshot::tests -- --nocapture — 7 passed
• ctrl-build --remote -- cargo check -p v1-compiler --bin claim_executor — passed
• generated_artifact_gate::main_wet — passed
• regen_stage0 --verify — regen_divergence_count=0
• cargo fmt --all --check and git diff --check — passed

Upstream #8116 is merged and this branch includes its repairs.

[gunbai-bot]

CI repair at 38254ac6bf: run 31439799302 reached ordinary batch 3 with 4,580 live-read requests, then spent about 2,697s in classification_production and crossed the 3,300,000ms floor budget by 237ms. The modeled request path rebuilt the same dependency adjacency inside every bound declaration's closure query. This commit adds dependency_closure_from_adjacency, prepares one LiveReadManifestPreparation before the manifest fold, and makes the per-request projection unable to access dependency edges for a cold reconstruction. The focused model control is green and proves legacy/prepared closure equality, declaration-grain discrimination, and row-order preservation. Fresh CI run: https://github.com/gunb-ai/gunbc/actions/runs/31447906159 — sent from wise-boar-328

[gunbai-bot]

Addressed review 51002 in a240b83b7f: PreparedFloorIdentityIndex now builds one multiplicity-preserving group_by index for selected requests and probes, and duplicate/missing/unexpected/serve lookups use those buckets instead of rescanning. The serve fold now prepends and reverses once at the ready boundary. all_floor_preparation_controls_hold executes green, including duplicate, missing, unexpected, provider-refusal, identity, and selected-order controls. — sent from wise-boar-328

[gunbai-bot]

Main integration is resolved at 3fa6ba0d97. Because main deliberately reverted #8110, this resolution does not restore either the G2 live consumer or the group_by primitive. The ordinary-floor adjacency optimization is therefore absent from the current diff: its live consumer was removed on main.

The actionable cost findings from review 51002 remain fixed without that reverted substrate. floor_preparation.dag now makes one identity scan per selected/probe population using the existing generic native Map<ScheduleWitnessEntry, ...> substrate, derives duplicate/missing/unexpected refusals by keyed membership, and serves probes by exact witness key. The success fold prepends and reverses once. all_floor_preparation_controls_hold evaluates true against the merged tree. — sent from wise-boar-328

[gunbai-bot]

Investigated CI run 31451041570. The failure is the ordinary floor discovery batch: the same 11 baseline witness reds (doc graph, roadmap, and stage0 partition-closure 5-second thread-CPU budget overruns), followed by the expected finalization refusal. Current main run 31450191212 fails with the same witness set and same stage0 budget pattern; the #8118 exact-subject/preparation witnesses are not implicated. This is therefore a pre-existing main-floor failure, not a regression in this PR, and I am not changing unrelated authority to mask it. — sent from wise-boar-328

[gunbai-bot]

CI run 31461241844 repeats the same 11 ordinary floor-discovery witness reds as the current main baseline (run 31450191212): the pre-existing orphan/doc-graph and stage0 budget witnesses fail, and finalization is refused because the ordinary worker fails. No PR-added witness is implicated, and the build/heal/regen checks remain green. I am not pushing an unrelated test/authority change; this requires a separate baseline-floor repair. — sent from wise-boar-328

[cursor]

Bugbot is not enabled for your account, so this pull request was not reviewed.

Enable Bugbot in the Cursor dashboard to get automatic reviews on future PRs.

[gunbai-bot]

Run 31476676378 has no source failure to fix: regen reached [floor-terminal] outcome=completed detail=walk terminal exit code 0, then the GitHub action wrapper timed out at its 15-minute limit. The ci job correctly fail-fasted on REGEN=failure. Build and generated-artifact healing passed. This is an infrastructure/job-timeout failure, not a PR regression; I am not changing source to mask it. — sent from wise-boar-328