route_fold needs a corpus differential (the three shadowed clauses are schedule-order-dependent)
Approved current scope — 2026-09-11
Add a narrow test-support observation of the runtime's actual route-fold decisions and compare it with an independent model using the existing seeded graph corpus. Cover sole-readership, intervening master reads, and one-master retention, including bounded constructed valid program shapes where compiler schedule ordering otherwise shadows the clauses. Preserve production routing decisions and PCM. Add the requested comment guard for plain_route_gains/node_kind coupling. Reuse existing corpus infrastructure; no new benchmark framework or exhaustive fixture campaign.
Own crates/graph/src/runtime.rs narrow test-support seam/comments and crates/graph/src/program/tests.rs (or existing directly relevant graph tests). No production route-fold algorithm, bind-error correction, policy, artifact pin or compiler edits. #221 must follow this issue's delivered coverage.
Gates: focused graph corpus and existing route-fold tests; each named safety clause must have a discriminating physical negative control or an explicit evidence-based scope ruling before PASS. Avoid self-comparison: oracle cannot call the same predicate it verifies. Maintain accepted/refused coverage and stable ordering. Separate negative and positive build targets. Strict focused Clippy, relevant graph tests and required PR/main CI; no timing claim. If existing corpus cannot exercise the frozen contract with one bounded extension, report before building a new harness.
Astra XHIGH scoping approved; user authorized execution. Astra LOW implements, Astra XHIGH independently verifies. Five attempts maximum; each coherent pass gets one adversarial verdict. Root checkpoints exact paths and pushes promptly when focused checks pass, before more implementation. At most two active issues: #220 and #162; #221 queued behind delivered #220. Isolated worktrees; no overlapping edits. Root/lane B owns artifact qualification and pinning. Preserve histories and failed evidence; never weaken gates or commit compiler-IR captures.
Record actual argv/environment/source/exits/logs externally; pause green for root checkpoint. Stop on first unexpected failure and report for bounded correction. Required exact-head PR and main qualification plus upstream GitHub synchronization precede closure. Remove clean delivered worktrees after preserving evidence/history. Historical model names below are superseded by user routing.
Historical issue body
Required follow-up F1 from strip Job 3's adversarial verification. The fold's three GREEN clauses (sole-readership of the last slot, the in-between master scan, one-master retain) are genuinely shadowed in every compilable session today — proven by five adversarially constructed shapes — but the shadowing rests on incidental deterministic-schedule facts (submix-destination edges sorting before output-destination routes; sidechain consumers scheduling before routes) that nothing pins. A scheduler-order change would make the sole-readership clause the only defense, with zero red tests.
Fix: an M1-style route_folds_over_program corpus differential driving the runtime's own route_fold clause code over the seeded random-graph corpus (house precedent: #208's scatter_redirects_over_program, which this same round demanded and landed). Must land BEFORE the next change in the route-fold/mixdown class. Also fold in a comment-level guard for the plain_route_gains/node_kind cascade coupling (F7 residual).
Astra LOW attempt 1 focused checkpoint
Added an appended cfg(test) observation seam calling actual route_fold, an
independent backward-writer oracle in the existing seeded cohort corpus, and
one accepted plus three bounded valid-program guard controls. Production
predicates are unchanged; coupling comment replacement preserves line count.
Focused route-fold fixture, seeded cohort corpus and formatting passed.
Evidence: /tmp/issue220-attempt1 actual argv/env/source/log/exit records.
Incidental lock ordering preserved externally and restored. Root checkpoints;
independent XHIGH validity/oracle review, three physical guard removals, focused
strict Clippy/broader graph tests and required PR/main qualification remain.
Attempt 1 review and bounded attempt 2
Astra XHIGH verified independent oracle, valid program controls and exact source
restoration. All three physical guard removals failed their intended assertion;
debug and release-unwind graph suites each passed71 tests. Strict Clippy then
failed items_after_test_module on the appended seam. Review also required the
stale no-red ledger to describe new evidence precisely. Root authorized moving
the seam before the existing test module and replacing the same25 comment lines.
Astra LOW applied only those changes, preserving production line numbers and all
predicates/oracles. Formatting, locked strict all-target/all-feature Clippy and
both focused route/corpus tests passed. Existing non-fatal Clippy configuration
warnings remain unchanged. Evidence: /tmp/issue220-attempt2. Candidate retention
has evidence for exclusive route readership; its master-equality conjunct remains
subsumed by whole-list association, without independent mutation credit. Earlier
full-suite/control receipts: /tmp/issue220-xhigh-review-9b00b42f/review.md.
Independent exact-correction review and required PR/main qualification remain.
Independent Astra XHIGH attempt 2 PASS
Reviewed032dc98d6b4d45761a8b0ccacea9eea5c6cfc70a. Seam relocation is byte-identical,
production locations/predicates and oracle unchanged;25-line ledger now states
exactly the evidence earned. Fresh formatting, strict Clippy and both focused
receipts passed; previous three physical controls and both71-test suites remain
applicable without repetition. Remote issue/spec/head parity and clean status
verified. Existing artifact product evidence remains applicable; required CI
must reproduce its pin. Review: /tmp/issue220-xhigh-review-032dc98d/review.md.
Only this evidence record follows; required PR/main qualification remains.
Delivered — 2026-09-11
PR #728 merged reviewed ddd99ee as c476f5b.
Required PR qualification34551686277 and exact-main qualification34552159920
PASSED, reproducing the unchanged artifact pin80abeec2. Independent oracle,
three physical controls, graph suites and strict Clippy evidence remain above.
This upstream record precedes verified GitHub closure. #221 may now start from
this delivered baseline; its implementation must preserve the new differential.
route_fold needs a corpus differential (the three shadowed clauses are schedule-order-dependent)
Approved current scope — 2026-09-11
Add a narrow test-support observation of the runtime's actual route-fold decisions and compare it with an independent model using the existing seeded graph corpus. Cover sole-readership, intervening master reads, and one-master retention, including bounded constructed valid program shapes where compiler schedule ordering otherwise shadows the clauses. Preserve production routing decisions and PCM. Add the requested comment guard for plain_route_gains/node_kind coupling. Reuse existing corpus infrastructure; no new benchmark framework or exhaustive fixture campaign.
Own crates/graph/src/runtime.rs narrow test-support seam/comments and crates/graph/src/program/tests.rs (or existing directly relevant graph tests). No production route-fold algorithm, bind-error correction, policy, artifact pin or compiler edits. #221 must follow this issue's delivered coverage.
Gates: focused graph corpus and existing route-fold tests; each named safety clause must have a discriminating physical negative control or an explicit evidence-based scope ruling before PASS. Avoid self-comparison: oracle cannot call the same predicate it verifies. Maintain accepted/refused coverage and stable ordering. Separate negative and positive build targets. Strict focused Clippy, relevant graph tests and required PR/main CI; no timing claim. If existing corpus cannot exercise the frozen contract with one bounded extension, report before building a new harness.
Astra XHIGH scoping approved; user authorized execution. Astra LOW implements, Astra XHIGH independently verifies. Five attempts maximum; each coherent pass gets one adversarial verdict. Root checkpoints exact paths and pushes promptly when focused checks pass, before more implementation. At most two active issues: #220 and #162; #221 queued behind delivered #220. Isolated worktrees; no overlapping edits. Root/lane B owns artifact qualification and pinning. Preserve histories and failed evidence; never weaken gates or commit compiler-IR captures.
Record actual argv/environment/source/exits/logs externally; pause green for root checkpoint. Stop on first unexpected failure and report for bounded correction. Required exact-head PR and main qualification plus upstream GitHub synchronization precede closure. Remove clean delivered worktrees after preserving evidence/history. Historical model names below are superseded by user routing.
Historical issue body
Required follow-up F1 from strip Job 3's adversarial verification. The fold's three GREEN clauses (sole-readership of the last slot, the in-between master scan, one-master retain) are genuinely shadowed in every compilable session today — proven by five adversarially constructed shapes — but the shadowing rests on incidental deterministic-schedule facts (submix-destination edges sorting before output-destination routes; sidechain consumers scheduling before routes) that nothing pins. A scheduler-order change would make the sole-readership clause the only defense, with zero red tests.
Fix: an M1-style
route_folds_over_programcorpus differential driving the runtime's ownroute_foldclause code over the seeded random-graph corpus (house precedent: #208's scatter_redirects_over_program, which this same round demanded and landed). Must land BEFORE the next change in the route-fold/mixdown class. Also fold in a comment-level guard for theplain_route_gains/node_kindcascade coupling (F7 residual).Astra LOW attempt 1 focused checkpoint
Added an appended cfg(test) observation seam calling actual route_fold, an
independent backward-writer oracle in the existing seeded cohort corpus, and
one accepted plus three bounded valid-program guard controls. Production
predicates are unchanged; coupling comment replacement preserves line count.
Focused route-fold fixture, seeded cohort corpus and formatting passed.
Evidence:
/tmp/issue220-attempt1actual argv/env/source/log/exit records.Incidental lock ordering preserved externally and restored. Root checkpoints;
independent XHIGH validity/oracle review, three physical guard removals, focused
strict Clippy/broader graph tests and required PR/main qualification remain.
Attempt 1 review and bounded attempt 2
Astra XHIGH verified independent oracle, valid program controls and exact source
restoration. All three physical guard removals failed their intended assertion;
debug and release-unwind graph suites each passed71 tests. Strict Clippy then
failed items_after_test_module on the appended seam. Review also required the
stale no-red ledger to describe new evidence precisely. Root authorized moving
the seam before the existing test module and replacing the same25 comment lines.
Astra LOW applied only those changes, preserving production line numbers and all
predicates/oracles. Formatting, locked strict all-target/all-feature Clippy and
both focused route/corpus tests passed. Existing non-fatal Clippy configuration
warnings remain unchanged. Evidence:
/tmp/issue220-attempt2. Candidate retentionhas evidence for exclusive route readership; its master-equality conjunct remains
subsumed by whole-list association, without independent mutation credit. Earlier
full-suite/control receipts:
/tmp/issue220-xhigh-review-9b00b42f/review.md.Independent exact-correction review and required PR/main qualification remain.
Independent Astra XHIGH attempt 2 PASS
Reviewed032dc98d6b4d45761a8b0ccacea9eea5c6cfc70a. Seam relocation is byte-identical,
production locations/predicates and oracle unchanged;25-line ledger now states
exactly the evidence earned. Fresh formatting, strict Clippy and both focused
receipts passed; previous three physical controls and both71-test suites remain
applicable without repetition. Remote issue/spec/head parity and clean status
verified. Existing artifact product evidence remains applicable; required CI
must reproduce its pin. Review:
/tmp/issue220-xhigh-review-032dc98d/review.md.Only this evidence record follows; required PR/main qualification remains.
Delivered — 2026-09-11
PR #728 merged reviewed ddd99ee as c476f5b.
Required PR qualification34551686277 and exact-main qualification34552159920
PASSED, reproducing the unchanged artifact pin80abeec2. Independent oracle,
three physical controls, graph suites and strict Clippy evidence remain above.
This upstream record precedes verified GitHub closure. #221 may now start from
this delivered baseline; its implementation must preserve the new differential.