ci(test-shards): the weekly shard-timings refresh is refused by the partitioner's own self-test — @objectstack/cli (1432 s) can be neither kept whole nor sliced under the density bound #29777

Activity

1. objectstack-fleet

   added

   domain:devx

   finding

   on Nov 3, 2026
2. objectstack-fleet commented on Nov 3, 2026

   

   objectstack-fleetBot

   on Nov 3, 2026

   ContributorAuthor

   More actions

   triage#1 · os-zhuang · session_01AavokzJ5DndAwitDXvKy4U

   Path: none

   Triage: first-touch audit · p2 · domain:devx kept · pm:queue · 2026-11-03T04:53Z

   ⛔ Not a claim.

   • Read fresh. Measured: the weekly shard-timings refresh (PR chore(ci): refresh the Test Core shard-timings dataset #28978) is refused by the partitioner's own self-test, because @objectstack/cli at 1432 s can be neither kept whole nor sliced under the density bound. Every weekly refresh stays red until the partitioner or the bound changes, and the next run is Monday 05:30Z. CI tooling, p2.

   ---

   Generated by Claude Code
3. objectstack-fleet

   added

   bugSomething isn't workingSomething isn't working

   pm:queue

   priority:p2Medium: important, M3Medium: important, M3

   pm:dispatched

   and removed

   finding

   pm:queue

   on Nov 3, 2026
4. objectstack-fleet

   assigned

   marchtian

   on Nov 3, 2026
5. objectstack-fleet commented on Nov 3, 2026

   

   objectstack-fleetBot

   on Nov 3, 2026

   ContributorAuthor

   More actions

   domain:devx#2 · marchtian · session_017ihL2gLghEqeDKTbDTC7Zd

   Claim: PM loop round 3
   Session: session_017ihL2gLghEqeDKTbDTC7Zd
   Account: marchtian
   Branch: claude/issue-29777-partitioner-slice-spread
   Worktree: objectstack-issue-29777
   Domain: domain:devx
   Seat: domain:devx#2
   File surface: scripts/partition-test-shards.mjs and its self-test; the shard count or density-bound declaration it reads, if that is where the fix lands; ⛔ not the generated timings dataset
   Container & model: M, mode:subagent, model: default tier, effort: high
   Clause-②: no
   Metadata-protocol: none
   Responsibility: the weekly shard-timings refresh is refused by the partitioner's own slice spread self-test, so it can never land
   Thread-read: 6595787705
   Premise: holds (PR #28978 head b5b4b790 red on Shard partitioner self-test only)
   Serial constraints cleared: PR #28978 is parked on this card (6595319265); no other open PR touches scripts/partition-test-shards.mjs at claim time (the dev re-checks)
6. objectstack-fleet commented on Nov 3, 2026

   

   objectstack-fleetBot

   on Nov 3, 2026

   ContributorAuthor

   More actions

   domain:devx#2 · marchtian · session_017ihL2gLghEqeDKTbDTC7Zd

   席位读数 — HOLD, a bound ruling owed · domain:devx#2 · base 97c2982c00 · 2026-11-03T05:40Z

   The dev reproduced PR #28978's red byte for byte (partition-test-shards --self-test on head b5b4b790's dataset: exit 1, slice spread), and stopped at the order's design-choice clause. No PR was opened and nothing was pushed. The seat's reading of the measurements:

   1. Pin 3b is right. On the refreshed dataset the full list would run @objectstack/cli whole: a 1432 s serial task plus the 270 s declaration closure, against about 500-570 s on the other seven shards.
   2. The refusal is structural. The CLI shrank (2697.45 to 1432.04 s; no file dominates it, the slowest is 21.67 s). It now sits below the 8-shard mean, so the whole split is balanced and the density cap equals the mean (1455.81 s). The declaration-closure charge (ci(test-core): build declarations on one CLI slice only — pin the CLI's declaration readers into slice 1/n #27018) then moves about 783 s onto the other seven shards, to about 1552.6 s at x2, x3 and x4 alike. On a balanced whole split the charge and the cap cannot both hold. On the committed dataset the cap (2697.45 s) has never bound.
   3. A second red sits behind it. Pin 3c derives x2 on the refreshed dataset (716 s against 736.45 s, a 20 s margin), while the committed dataset refuses x2. So the slice-count map and the dataset can only land together.

   Options measured for which bound yields (prototypes in a scratch tree only):

   • B: the cap gains charge/(N−1) only when the run carries the declaration slice (1567.67 s at 8 shards). x2 and x4 fit, and the cap still binds.
   • A: floor the cap at the 1868 s whole-leg figure. x2, densest 1553.36 s.
   • C: keep the cap and accept the CLI whole. That costs about 7-11 min serial on every full run.
   • D: 9 shards (ci.yml, Tier H). This is not a fix by itself.
   • E: drop the closure charge. It fits by 0.02 s and undoes ci(test-core): build declarations on one CLI slice only — pin the CLI's declaration readers into slice 1/n #27018.

   The dev recommends B, landed first (it is inert on today's dataset), followed by one lane PR that carries the generated dataset byte for byte with the map 4→2 (the #26186 shape), which replaces #28978. This changes what a bound means, so the seat sends it to the maintainer instead of ruling. #28978 stays parked, and this card waits on that ruling.
7. objectstack-fleet commented on Nov 3, 2026

   

   objectstack-fleetBot

   on Nov 3, 2026

   ContributorAuthor

   More actions

   domain:devx#2 · marchtian · session_017ihL2gLghEqeDKTbDTC7Zd

   Maintainer ruling recorded · domain:devx#2 · 2026-11-03T06:22Z

   The maintainer, on the project's decision card, question 「分片密度上限和声明闭包计费,放宽哪一个?」, picked option 「B:上限加闭包分摊(推荐)」 (a platform-recorded pick, so the option label is the maintainer's word). This answers the seat's HOLD 6596381272.

   What the seat executes:

   1. B lands first. In planShards, when the WITHIN-CAP check reads a candidate split that carries the declaration reader slice, the cap gains DECLARATION_CLOSURE_CHARGE_SECONDS / (shardCount − 1). The self-test's hard constraint that the committed dataset's full list stays within the cap is scored against that same adjusted cap. The pin "cap = the whole split's densest shard" and the capFixture positive control stay as they are. On today's committed dataset the output must stay byte-identical; the bins read 1600.11/1597.48/… as now.
   2. Then one lane PR carries the refreshed dataset byte for byte, with FILE_SHARDED_PACKAGES cli 4→2 and PREVIOUS_FILE_SHARDED_PACKAGES = 4 (the ci(test-shards): refresh the Test Core dataset and cut the CLI into 4 slices on plain-sum placement, and pin the whole-package leg #26186 shape). It replaces chore(ci): refresh the Test Core shard-timings dataset #28978, which then closes. ci.yml's DTS_READER_SLICE and the cli vitest config's DECLARATION_READER_SLICE do not change.

   Measured on the refreshed dataset at 8 shards: the B cap is 1567.67 s against a densest x2 shard of 1553.36 s, a 14 s margin. Pin 3c's lack of hysteresis (x2 holds by only 20 s) is outside this ruling and goes to its own finding.
8. objectstack-fleet commented on Nov 3, 2026

   

   objectstack-fleetBot

   on Nov 3, 2026

   ContributorAuthor

   More actions

   domain:devx#2 · marchtian · session_017ihL2gLghEqeDKTbDTC7Zd
   os-dev-report

   Round 2: slice 1 of ruling record 6596963071 (option B). Draft PR #29838, head e9569e920e. Measured on objectstack-ai/objectstack, merge base b0ef25f524.

   {
     "issue": 29777,
     "status": "done",
     "branch": "claude/issue-29777-partitioner-slice-spread",
     "pr": "https://github.com/objectstack-ai/objectstack/pull/29838",
     "head_sha": "e9569e920e",
     "session": "session_017ihL2gLghEqeDKTbDTC7Zd",
     "premise_still_valid": true,
     "summary": "planShards() judges, and repairs, a candidate split that carries the declaration-reader slice against densityCap + DECLARATION_CLOSURE_CHARGE_SECONDS / (shardCount - 1) (new closureAllowance(); 111.86 s at 8 shards); a split without it keeps the bare cap. The self-test hard constraint is scored against the same adjusted cap; the cap-definition pin and the capFixture controls are unchanged. 4 new pins on the refresh's frozen weights (79 packages from b5b4b790, inside the test): x2 taken within 1567.67 s (densest 1553.36 s), not carried allows 0 s, a cap short of the adjusted one still refuses by name. Today's committed dataset splits byte-identically.",
     "tests": [
       "partition-test-shards --self-test, committed dataset at e9569e920e: exit 0; verdict line byte-identical to base",
       "byte-identity base b0ef25f524 vs head: main() --shard 1/8..8/8 --exclude @objectstack/dogfood on a full-list document, stdout and stderr byte-identical on all 8; plan JSON items, bins 1600.11/1597.48/1597.14/1599.09/1598.01/1597.78/1599.48/1597.04 and cap 2697.45 identical; the one difference is the added decision key allowance 111.85714285714286",
       "PR 28978 dataset (b5b4b790), scratch tree: base red on 7 checks (first: slice spread), head red on 1 (pin 3c 'Lower it to 3', the map swap slice 2 carries); with slice 2's map applied in scratch (cli 2, previous 4): exit 0, CLI x2, bins max 1553.36 within 1455.81 + 111.86 allowance, whole legs within 1868",
       "ablation, scratch tree: closureAllowance mutated to return 0 (marker grep-confirmed): exit 1; collected, the 3 allowance-dependent cases red, the not-carried case green; restored by byte comparison, restored exit 0",
       "importer self-tests exit 0: nightly-tiers, check-test-suite-ceilings, measure-test-shard-timings, ci/select-shard-timings-run, ci-shard-matrix, check-test-completeness"
     ],
     "gates": "dispatch-gates --commands at e9569e920e: 31 commands, all run with exit codes captured before any pipe after pnpm install --frozen-lockfile, all 31 exit 0 (pnpm check:pm-dispatch-gates fast tier 1945 cases, exit 0). --ran: 31 derived, 31 run, 0 NOT-MEASURED, a DERIVED zero (every line carries its exit code). First pass before install: 6 exit 3 PREREQUISITE NOT MET (yaml missing) + 1 environment finding (@changesets/cli missing), all green after install. NOT MEASURED locally, CI's: the Plan Test Core shards step and the type-check lanes. PR check-runs at report time: 10 completed, 0 failed, 15 in_progress.",
     "dogfood_battery": "n/a: no ! in the title and no (narrowing) arm",
     "line_budget": "+168 / -15, 1 file; under the 3000 human-merge threshold",
     "deviations": [
       "Round 1 removed a scratch worktree with --force; it held only my own intentional ablation mutation. This round's scratch tree is the same case, removed after its last read.",
       "Byte-identity holds for the split and every shard's output; the slice-decision record gains one key (allowance), named above.",
       "The self-test success line was changed so it names the allowance when bins exceed the bare cap; its text today is byte-identical, and without the change it would read false once slice 2 lands.",
       "PR body ends with the line the order specified; the session URL sits in the body prose."
     ],
     "files_changed": ["scripts/partition-test-shards.mjs"],
     "claims": "Claim 6596162760 (this branch, this session); no second claim; card assignee untouched; PR has no assignees and no reviewers.",
     "measurements": "Round 1's per-candidate table stands. Under B at 8 shards on the refreshed dataset: cap 1455.81, allowance 111.86, judged 1567.67; x2 densest 1553.36 (14.31 s margin). Committed dataset unchanged: cap 2697.45, bins as above.",
     "mcp_calls": "0",
     "api_writes": "2: pr_create via the relay (run 41592767706, POST pulls plus the skip-changeset label leg; read-back 7224 bytes sent and stored, identical) and this comment via post-stamped.mjs. git push x2 of the own branch (not REST).",
     "open_questions": [],
     "security_findings": "none",
     "out_of_scope_findings": [
       "class: reading, the ruling files it separately · pin 3c has no hysteresis: refreshed x2 holds by 20 s (716 vs 736.45); carrier: domain:devx seat",
       "class: reading (not filed) · the committed layout re-weighed with the refreshed weights carries 2161 s of whole packages on one shard, past the 1868 s whole-leg cap; carrier: domain:devx seat"
     ]
   }

   ---

   Generated by Claude Code
9. objectstack-fleet

   closed this as completedin #29838on Nov 3, 2026
10. objectstack-fleet commented on Nov 3, 2026

    

    objectstack-fleetBot

    on Nov 3, 2026

    ContributorAuthor

    More actions

    domain:devx#2 · marchtian · session_017ihL2gLghEqeDKTbDTC7Zd

    Landed, both slices · domain:devx#2 · 2026-11-03T08:28Z

    • Slice 1: PR ci(test-shards): the density cap carries the declaration closure's own displacement when a split carries the reader slice #29838 merged as 3cdb162fa5. The density cap now carries DECLARATION_CLOSURE_CHARGE_SECONDS / (N−1) when a split carries the declaration slice. The output on the committed dataset was unchanged.
    • Slice 2: lane PR ci(test-shards): carry the 2026-11-02 shard-timings refresh and slice the CLI in 2 #29856 merged as 4b71a82b76. It carries the 2026-11-02 refresh byte for byte, with FILE_SHARDED_PACKAGES cli 4→2 and PREVIOUS_FILE_SHARDED_PACKAGES = {cli: 4}. The self-test reads the CLI at x2 with a densest shard of 1553.36 s, within 1567.67 s. PR chore(ci): refresh the Test Core shard-timings dataset #28978 closes as superseded.
    • Pin 3c's 20 s margin is finding ci(test-shards): the slice-count pin has no hysteresis, so a 20 s drift flips the CLI map and turns the weekly refresh red again #29826.

    The seat releases the card: pm:dispatched comes off, and the worktree is removed.
11. objectstack-fleet

    removed

    pm:dispatched

    on Nov 3, 2026