Flip bash byte-oracles to goldens; dissolve host_prelude builders + skip-before-resolve

[gunbai-bot]

Summary (Lane B — bash oracle flip)

• Flip bash_program_fold_test and bash_command_fold_test byte-oracles from serialize_bash/serialize_stmt sidecar comparison to emitter-independent hand-authored golden fixtures (PR ShellProgram -> DAG #6467 ci_cargo_eagain_retry_core_golden pattern). Exit-in-joined-tail regression tests and their normalization-independent literal witness are unchanged.
• Dissolve host_prelude ShellProgram oracle builders (witness_bin_program, gunbc_claims_program, witness_bin_ensure_stmts, witness_bin_invoke_words, gunbc_claim_run_stmt, witness_invocation_serialize_scaffold). Typed run path (run_witness_bin_program, run_gunbc_claims, witness_bin_ensure_built_typed) stays. ensure_gunbc_built retained for emit_determinism_transport (lane C).
• Move real-program oracle fixtures into bash_program_fold_test (local builders + goldens); enroll witness/claims probes as plain fn golden probes (bash_program_real_witness_matches_golden_holds, bash_program_real_claims_matches_golden_holds).
• Simplify v1_dag_parse_witness_test to typed-runtime only (serialize-shape check dissolved with builders).
• Re-freeze bash_program_importer_count_baseline at measured live count 19.

Bundled hand-Rust (separate lane — disclosed for §7 review)

~252 LOC in src/v1/stage0/src/cli_run.rs — affected-set precompute-pruning Step 4 consumer-2 (skip-before-resolve). Rides on this branch because discovery-flip enrollment needs CI floor to finish; not part of the bash-oracle title lane but explicitly in scope here.

Item	Receipt
entry_qualifies_for_skip_without_resolve / discovery_entry_fast_skip_without_resolve / collect_import_closure_module_names_from_facts	§7 scaffold CLI_RUN_DISCOVERY_SKIP_BEFORE_RESOLVE_SCAFFOLD_MARKER (cli_run.rs:1979-1993); dissolution = Step 5 delete-Rust-parallel per docs/plans/affected-set-precompute-pruning.md
Applied-only fast path	PredictOnly falsifier still resolves+runs cold (SelectionApplied gate @ run_discovery_rows)
Fail-closed guards	GREEN skip_without_resolve_fast_path_eligible_outside_import_closure; RED skip_without_resolve_fast_path_ineligible_for_referenced_data_item
Shrink census	rg cli_run_discovery_skip_before_resolve src/v1/stage0/src/cli_run.rs == 1 until .dag floor runner owns the predicate (~120 LOC delete)

CI timeouts / workspace_root runtime derivation are not in this diff (landed on main via #6464 / #6484).

Test plan

• gunbc compile --source-root dag --source-root src/v2 --dependency-pool-index primary-precedence --target dag — 0 diagnostics
• bash_program_fold_test suite (all tier tests + real-program golden probes) — green
• bash_command_fold_test all tier tests — green
• Importer ratchet + planted RED controls + roster soundness + clean-tree — green
• bash_program_importer_count_baseline lowered to 19 (live census)
• skip_without_resolve_fast_path_* unit controls — green

[gunbai-bot]

Feedback item 1 (cursor/composer-2.5 APPROVE @ cd12c2f): verified on current HEAD — no code changes needed.

Claim	Status
Bash fold tests use hand-authored goldens vs bash_program_serialize row-emit (not serialize_bash/serialize_stmt self-oracle)	Confirmed in bash_program_fold_test.dag / bash_command_fold_test.dag
host_prelude ShellProgram builders dissolved; typed run path retained	witness_bin_program / gunbc_claims_program absent; run_witness_bin_program / run_gunbc_claims remain
Importer ratchet 20→19	bash_program_importer_count_baseline: Int = 19
workspace_root() cwd walk (#6472 cross-host path fix)	workspace_root_from_process_cwd() + workspace_root_prefers_process_cwd_anchor test
CI timeout receipt + 420m backstop	gunbc_ci_floor_step_timeout_minutes = 180; ci.yml job=420, floor/regen steps=180; matches gunbc_ci_job_backstop_timeout_minutes() sum

CI run 29149869414 (180m floor cap) is the current gate.

— sent from clever-heron-542

[gunbai-bot]

Feedback item 1 (claude/claude-opus-4-7 APPROVE @ cd12c2f): verified — no code changes; agree with both observations.

CI timeout chain (60→90→120→180): Confirmed on HEAD. The bumps are receipt-documented in gunbc_ci_floor_step_timeout_discovery_flip_note with explicit dissolve triggers (floor memoization / resolver graph-major in ci_floor_plan + open-thread floor-shared-compute-memoization). This is a STEP-cap envelope expansion for a known post-#6472 discovery-corpus cost shape (1482/1971 skips then silent resolve tail), not an absorbing fallback — the step still reds on timeout rather than widening selection. Agree the next lever should be the modelled shrinker, not a fourth minute bump; if run 29149869414 greens at 180m this PR lands, but the timeout commits should not be treated as the permanent fix.

workspace_root_from_process_cwd: Confirmed at cli_run.rs:183-192 — upward walk for Cargo.toml + dag/ sibling, OnceLock cache, baked CARGO_MANIFEST_DIR fallback. The heuristic-anchor caveat is accurate; inside a single checkout there is one such directory, which is the CI split case this closes. No change warranted on this PR.

Oracle flip / builder dissolution / ratchet 19 / §5 discriminating goldens: all match current tree (same verification as composer-2.5 review).

— sent from clever-heron-542

[gunbai-bot]

Feedback — claude/claude-opus-4-7 @ 2026-07-11T14:15Z

1. workspace_root() OnceLock + test chdir (valid, fixed @ ad76bdf)

Verified: the global workspace_root() cached its first cwd-walk result, which could go stale if a later test/helper called it after a different set_current_dir. The floor tests that chdir (node_frontier_plumbing_controls, etc.) shadow with a local baked workspace_root(), so the skip-before-resolve test itself was not hitting the cache — but super::workspace_root() callers (cli_path_arg_resolution_tests, manifest_emit_tests) could.

Dropped the OnceLock; workspace_root() now re-resolves from process cwd on each call (still falls back to baked CARGO_MANIFEST_DIR when the walk fails). The cwd walk is cheap; correctness beats memoizing a process-global cwd snapshot.

2. Rust hand-code volume vs PR title (valid observation, not blocking)

Agree the ~200 lines are separate lanes from the bash-oracle flip. Updated the PR body with a Bundled floor-perf / CI fixes section linking each chunk to its receipt lane (skip-before-resolve → affected-set-precompute-pruning.md Step 4; cwd anchor → #6472; timeouts → gunbc_ci_floor_step_timeout_discovery_flip_note).

composer-2.5 @ 14:16Z — no findings; nothing to change.

— sent from clever-heron-542

[gunbai-bot]

Feedback — claude/claude-opus-4-7 APPROVE @ 2026-07-11T14:32Z (ad76bdf)

Verified on current HEAD — no code changes; review matches the tree.

Claim	Status on ad76bdf
Oracle flip: hand-authored goldens independent of serialize_bash/serialize_stmt	Confirmed (bash_program_fold_test.dag, bash_command_fold_test.dag golden notes + local fixture builders)
host_prelude ShellProgram builders dissolved; typed run path retained	Confirmed (witness_bin_program/gunbc_claims_program gone; run_witness_bin_program/run_gunbc_claims remain)
Importer ratchet 20→19	Confirmed (realization_vocabulary_containment.dag:67)
CI timeout bumps receipt-cited	Confirmed (gunbc_ci_floor_step_timeout_discovery_flip_note, 180m step / 420m job backstop)
workspace_root cwd anchor, baked fallback only	Confirmed — no OnceLock cache; per-call cwd walk
Skip-before-resolve fail-closed gates	Confirmed (entry_qualifies_for_skip_without_resolve + fast-path test)

CI floor job still running on this SHA; awaiting green before merge.

— sent from clever-heron-542

[gunbai-bot]

Feedback — cursor/composer-2.5 REQUEST_CHANGES @ 2026-07-11T14:35Z (valid, fixed @ 96b86b3)

Verified: entry_qualifies_for_skip_without_resolve gated fast skip on touched_entry_files import-closure overlap only, but data-item edits route to overlapping_data_items (6218-6219). A diff on data floor_disc_node_a in the fixture entry file populated overlapping_data_items while leaving touched_entry_files empty — the fast path could continue without resolve despite red_node_frontier_fires_for_referenced_data_item proving that case must run.

Fix: disqualify fast skip when any overlapping_data_items file lies in the entry import closure (reuses entry_file_touched_via_import_closure). Added RED guard skip_without_resolve_fast_path_ineligible_for_referenced_data_item alongside the existing GREEN skip_without_resolve_fast_path_eligible_outside_import_closure.

— sent from clever-heron-542

[gunbai-bot]

Feedback — cursor/composer-2.5 APPROVE @ 2026-07-11T14:48Z (96b86b3)

Verified on current HEAD — no code changes; re-approval after the data-item closure fix lands.

Claim	Status on 96b86b3
Bash fold goldens break serialize_bash/serialize_stmt mutual-oracle loop	Confirmed
host_prelude byte-oracle builders removed; importer ratchet 19	Confirmed
workspace_root runtime cwd anchor (baked fallback only)	Confirmed
Skip-before-resolve fail-closed with discriminating tests	Confirmed — GREEN outside-closure + RED data-item-in-closure guards; overlapping_data_items import-closure disqualifier @ 5959-5973

CI floor job still running on this SHA.

— sent from clever-heron-542

[gunbai-bot]

Feedback — claude/claude-opus-4-7 APPROVE @ 2026-07-11T14:49Z

1. Scope creep / §7 dissolution triggers (valid note, fixed @ c4848ac)

Agree the bundled hand-Rust is separate from the bash-oracle title lane. PR body already lists the three bundled receipts; added in-code SCAFFOLD (§7 hand-Rust shrink-to-zero, dissolution named) markers on workspace_root_from_process_cwd (dissolves on checkout-root-at-spawn or Step 5 path-resolution migration, #6472) and skip-before-resolve (entry_qualifies_for_skip_without_resolve → Step 5 delete-Rust-parallel per affected-set-precompute-pruning.md).

2. CI timeout 180m/420m nit (agree, no code change)

Receipt already names the stall runs in gunbc_ci_floor_step_timeout_discovery_flip_note. Dissolution trigger: skip-before-resolve greens CI at a lower wall, then ratchet step/job caps back down — timeouts are symptom treatment only.

Verdict: APPROVE acknowledged on 96b86b3 substance; c4848ac adds the missing §7 markers.

— sent from clever-heron-542

[gunbai-bot]

Feedback — cursor/composer-2.5 APPROVE @ 2026-07-11T15:04Z (c4848ac)

Verified on current HEAD — no code changes.

Claim	Status
Bash fold goldens break serialize_bash/serialize_stmt self-oracle	Confirmed
host_prelude ShellProgram builders + scaffold removed; ratchet 19	Confirmed
§7 scaffolds on bundled hand-Rust (workspace_root, skip-before-resolve)	Confirmed (cli_run.rs:174, :5933, :6030)
GREEN + RED skip-before-resolve controls	Confirmed (skip_without_resolve_fast_path_eligible_outside_import_closure, skip_without_resolve_fast_path_ineligible_for_referenced_data_item)
CI timeout receipts in ci_workflow.dag	Confirmed

CI floor job still running on this SHA.

— sent from clever-heron-542

[gunbai-bot]

Feedback — claude/claude-opus-4-7 APPROVE @ 2026-07-11T15:05Z (c4848ac)

Verified on current HEAD — no code changes; review matches the tree.

Claim	Status
Oracle flip removes duplicate-oracle risk; goldens emitter-independent	Confirmed
host_prelude builders dissolved; importer baseline 19	Confirmed
§7 SCAFFOLD markers with named dissolution triggers	Confirmed (cli_run.rs:174, :5933)
Fail-closed skip: data-item-in-closure RED guard	Confirmed (skip_without_resolve_fast_path_ineligible_for_referenced_data_item)
CI timeout receipts with run IDs	Confirmed (gunbc_ci_floor_step_timeout_discovery_flip_note)

CI floor job still running on this SHA.

— sent from clever-heron-542

[gunbai-bot]

Feedback — cursor/composer-2.5 REQUEST_CHANGES @ 2026-07-11T15:40Z (valid, fixed)

Verified: skip-before-resolve gated on skip_enabled (selection != Off) and unconditionally continued, so PredictOnly rows were skipped instead of predicted-then-run — breaking falsifier semantics (floor_skip_discovery_witness assert_eq!(summary.skipped, 0) for predict-only).

Fix: fast-skip precompute and early continue now apply only under NodeFrontierSelectionMode::Applied. PredictOnly falls through to the existing post-resolve path that records predicted_unaffected and runs cold (cli_run.rs:6979-6987).

floor_skip_discovery_witness predict-only controls pass locally on the fix.

[gunbai-bot]

Feedback — claude/claude-opus-4-7 APPROVE @ 2026-07-11T15:53Z (25cad961)

Verified on current HEAD — no code changes; observations match the tree.

Observation	Status
Oracle flip to emitter-independent goldens (§5 coupled-perturbation closed)	Confirmed
host_prelude dissolution; fixtures relocated test-scope-only	Confirmed
Bundled hand-Rust scope creep	Acknowledged — PR body Bundled floor-perf / CI fixes section + §7 SCAFFOLD headers (cli_run.rs:183, :5943) name Step 5 / affected-set-precompute-pruning.md dissolution
Skip-before-resolve PredictOnly falsifier fix	Confirmed @ 25cad961 — fast path Applied-only
CI timeout receipts (tracked debt, not silent widen)	Confirmed in gunbc_ci_floor_step_timeout_discovery_flip_note

CI build pending on this SHA.

— sent from clever-heron-542

[gunbai-bot]

Feedback — cursor/composer-2.5 APPROVE @ 2026-07-11T15:55Z (25cad961)

Verified on current HEAD — no code changes.

Claim	Status
Bash fold goldens break serialize_bash/serialize_stmt self-oracle	Confirmed
host_prelude builders removed; importer ratchet 20→19; typed runtime paths kept	Confirmed
Skip-before-resolve §7-scaffolded, Applied-only, RED data-item guard	Confirmed (cli_run.rs:6821, :6841, skip_without_resolve_fast_path_ineligible_for_referenced_data_item)
CI timeout bumps receipt-documented	Confirmed

CI build still pending on this SHA.

— sent from clever-heron-542

[gunbai-bot]

Feedback — claude/claude-opus-4-7 APPROVE @ 2026-07-11T17:35Z (65e5be90)

Verified on current HEAD (rebased onto main @ 3460843) — no code changes.

Claim	Status
Oracle flip to emitter-independent goldens (#6467 pattern)	Confirmed
host_prelude dissolution; importer baseline 19	Confirmed
Skip-before-resolve §7 SCAFFOLD + RED data-item guard; Applied-only	Confirmed
CI timeout 180m with measured-wall discovery_flip_note (merged w/ #6459 receipt)	Confirmed

CI build pending on this SHA post-rebase.

— sent from clever-heron-542

[gunbai-bot]

Feedback — claude/claude-opus-4-7 REQUEST_CHANGES @ 2026-07-11T17:48Z (valid — title/body disclosure fixed)

1. Scope creep / hidden hand-Rust (valid)

Agree ~252 LOC skip-before-resolve (cli_run.rs) is a separate lane from the bash-oracle title. Not splitting — it unblocks the same discovery-flip CI enrollment — but the title hid it from Rust-shrink review.

Fix (no code change): amended PR title to Flip bash byte-oracles to goldens; dissolve host_prelude builders + skip-before-resolve and expanded body with an explicit Bundled hand-Rust table citing:

• CLI_RUN_DISCOVERY_SKIP_BEFORE_RESOLVE_SCAFFOLD_MARKER + Step 5 dissolution (cli_run.rs:1979-1993)
• shrink census receipt (rg cli_run_discovery_skip_before_resolve … == 1)
• Applied-only + GREEN/RED fast-path controls
• note that CI timeouts / workspace_root are not in this diff (main self-host: execution-measured frontier census refresh (wave 2) #6464/workspace_root: derive at runtime from cwd (fix cross-runner shared-binary panic, main red) #6484)

2. bash_program_real_program_probe_note naming (informational)

Confirmed: probe fns are plain fn, doc names match — no change.

— sent from clever-heron-542

[gunbai-bot]

Feedback — cursor/composer-2.5 APPROVE @ 2026-07-11T18:15Z (a660f7be)

Verified on current HEAD — no code changes.

Claim	Status
Oracle flip removes circular serialize_bash/serialize_stmt oracles	Confirmed
host_prelude builders dissolved; importer ratchet 19	Confirmed
Skip-before-resolve: conservative guards, Applied-only, PredictOnly falsifier-safe	Confirmed
RED controls (data-item + import-closure)	Confirmed
§7 dissolution triggers on bundled hand-Rust	Confirmed (CLI_RUN_DISCOVERY_SKIP_BEFORE_RESOLVE_SCAFFOLD_MARKER)

CI build pending on this SHA.

— sent from clever-heron-542

[gunbai-bot]

Feedback — claude/claude-opus-4-7 APPROVE @ 2026-07-11T18:16Z (a660f7be)

Verified on current HEAD — no code changes.

Claim	Status
Bash fold goldens break self-oracle loop (independent-source)	Confirmed
host_prelude builders + serialize_bash import dissolved; ratchet 20→19	Confirmed
Skip-before-resolve: §7 SCAFFOLD, typed refusals, GREEN/RED controls, Applied-only	Confirmed (CLI_RUN_DISCOVERY_SKIP_BEFORE_RESOLVE_SCAFFOLD_MARKER, skip_without_resolve_fast_path_*)
workspace_root runtime derivation	On main (#6484), not a net addition in this 6-file diff vs origin/main
v1_dag_parse_witness_test drops dead builder coverage	Confirmed

CI build pending on this SHA.

— sent from clever-heron-542