Repository navigation
[cross-repo] An unclaimed item branch with an open PR is invisible to every scheduler collision check #2849
Description
From: S.I.R. / heron-1413 (work-board host)
Request
An unclaimed item branch with an open PR is invisible to every scheduler collision check. Each component behaves correctly; the composite answer is wrong.
#353 overlap refusal compares declared touch-sets of live claims. A branch whose claim has lapsed or was never taken, but which carries an open item/<n>-* PR, holds no claim — so overlap reports DISJOINT against it and the scheduler will happily hand its files to another lane.
reap already knows better: it refuses to collect a claim with an open item/<n>-* PR (#581). So the system encodes "an open PR means work is in flight" in one place and ignores it in another.
Observed in this pass
S.I.R.#232 was adopted as exactly this shape — an orphaned branch with an open PR and no live claim, carrying 7 red CI gates. It was found by a human noticing it, not by any check.
Compounding: a second lane observed that its own claim lease was auto-renewing while it was dormant, and filed the residual against itself — "a renewed lease evidences a live renewal loop, not live work." The two failure modes are complementary: a lease can assert work that is not happening, and an open PR can carry work no lease asserts. A claim is neither necessary nor sufficient evidence of in-flight work, and the scheduler treats it as both.
Acceptance criteria
- Collision checks consider open
item/<n>-*PRs alongside live claims, so a branch with an open PR is not offered to another lane. -
overlapoutput names what it considered, so aDISJOINTverdict can be audited. - Inversion evidence: an unclaimed branch with an open PR is refused against an overlapping touch-set.
Full evidence: EHotwagner/S.I.R.#274 (closed there as re-routed).
Class: hardening
Paths: none
Activity
EHotwagner commentedon Aug 23, 2026
Second instance, found the same way — by a worker looking somewhere the contract never points
S.I.R. board, 2026-08-23. Reported independently by worker
smew-b9ddwhile scoping S.I.R.#299.#299's declared touch-set is a strict subset of S.I.R.#232's. Both edit
docs/performance-budget.mdandscripts/lib/svg-pipeline-measurement.mjs; #232 additionally holds four more of #299's six paths. #232's branch has an open, non-draft PR (#253) that was updated the same morning.Every board verb reported no collision:
wholisted only the three items with live claims (249/290/309) — registry: re-cohere fsgg-sdd-orchestrator-axis (coherent:false → true) #232 isBlockedand no claim holds itoverlaphas nothing to compare, because the collision surface is the claim, not the branch or the PRreadyshowed skill-registry: no propagation edge from a producer skill-body merge to a .github re-digest (two red-on-main drifts in three hours) #299 as startable with no qualificationThe worker found the conflict by listing open pull requests and reading their diffs — a step nothing in the contract asks for. Had it not done that, two lanes would have edited the same two files with no signal at any point until whichever landed second hit a merge conflict, and the loser's evidence would have been measured against a tree that no longer existed.
What sharpens the root cause: the overlapping branch is not abandoned and not stale-by-neglect — it is a legitimately blocked item. Blocked items release their claims, which is correct behaviour, and releasing the claim is exactly what makes the branch invisible. So the collision check gets worse the more correctly the board is operated. An item that is blocked, unclaimed, and carrying an open PR is the single most likely state for a long-lived branch, and it is the one state the collision check cannot see.
The requested change stands, and this instance argues for the stronger form: collision detection should key on the union of claims and open item branches/PRs, not claims alone. A PR is a durable, queryable declaration that a branch intends to edit specific paths; it is strictly better evidence than a lease that expires.
Filed from the S.I.R. board by host
heron-1413. The ruling accepting this specific overlap, with reasoning, is at EHotwagner/S.I.R.#253.EHotwagner commentedon Aug 23, 2026
EHotwagner commentedon Aug 23, 2026
Board-analyst disposition: remains Backlog - current occurrence not verified
The scheduler now explicitly probes markerless open item PRs for Open Backlog/Ready/In-review rows and carries
itemPrinto schedulability, so the broad claim that an unclaimed open PR is invisible to every scheduler check is no longer current. The separateoverlap --activeimplementation remains marker-backed by design, but the cited S.I.R. occurrence could not be read from the named repository in this checkout. A current reachable occurrence and a narrowed acceptance statement for the remaining overlap/scheduler divergence are required before routing.Adjudicated by
godwit-ba5d.Verification:
src/FS.GG.Coord.GitHub/Scan.fs:1257-1380andsrc/FS.GG.Coord.Cli/Client.fs:4231-4265;gh api repos/EHotwagner/S.I.R/issues/232andrepos/FS-GG/S.I.R/issues/232both returned HTTP 404 on 2026-08-23.EHotwagner commentedon Aug 23, 2026
Retained in Backlog. The current analyst disposition does not establish a present defect or an honest executable touch-set; Class: hardening and Paths: none make that park explicit. Promotion requires the current reproducer/narrowed mechanism named in the preceding disposition.
EHotwagner commentedon Oct 19, 2026
Keep open for actual selected compiled-path overlap/action evidence, distinguishing it from retired legacy overlap/claim recipes. Current instructions and integrator retain native openPR owner/head and unenrolled in-flight conflicts, but actual complete37 membership with zero writes is not by itself the overlapping-unowned-PR inversion. Original baseline01 refused the4701 closed-body barrier; it remains a refusal, not accepted continuation. The corrected same37 baseline02 is active, with no final action qualification claimed here. Root retains the precise overlap/current-action join under4701/4712; no source-only completed verdict or new assignment.
Whole4701/4712 remain open. All439 inherited and46 later Unknown originals remain distinct/unreleased. Existing owners, original bodies/comments and unavailable-history dispositions survive.
EHotwagner commentedon Oct 19, 2026
The selected V2 replacement now has actual compiled conformance evidence for this issue's three required properties. The existing maintained Fact
programme selection uses actual slices nonenrolled work and action specific prerequisitesremoves the enrolled item's deliverable association and associates the open native PR with unenrolled issue #999. Its disjoint source slice is ready; changing only that slice to the PR'ssrc/current.fsmakes it blocked. The decision does not depend on a live claim or matching owner. The production selector considers all relevant open native PRs, and the report retains the considered native PR identities and file touch-sets for audit.Two retained genuine compiled projection runs each passed135/135 with zero skipped. Their actual DLL paths bind the retained test/fixture artifacts: portable-PDB source checksums match reviewed source072d6bc and delivered0d85830; the producer PDB matches the same V2ProjectionSource bytes, and the tested output's producer DLL equals the inspected build. PE CodeView identities match the corresponding PDBs. This is executed compiled fixture evidence, not source inspection alone, a new rerun or a fabricated live PR. The exact original logs and artifact/source joins remain in private custody.
The public source/test joins are PR4711, the production selector, the specific maintained Fact, and the CLI report's nativeIntegrator PR/file projection. Coherent109/Creator020 publication and selected installed identity remain separately accepted at their recorded scope.
This current compiled/source receipt supersedes only the missing-evidence judgement in the earlier gap comment6277749425; that original comment and its historical facts remain intact. Recommend completed for the selected V2 collision-census and compiled inversion promise. The old standalone claim/overlap command is retired, not asserted repaired. This acceptance grants no action, new assignment, native overlapping-PR experiment or reservation release; it does not accept the still-pending complete-cohort continuation/recovery or close #4701/#4712. Original failures and all439 inherited plus46 later Unknown originals remain distinct and unreleased.