Skip to content

99 seals orphaned by the naming convention change, 81 with a current twin #2477

Closed as not planned

Description

@gHashTag

check_seal_coverage reports 121 stale seals. Measured, most are not stale work — they are duplicates left behind when the seal filename convention changed.

The mechanism

t27c seal --save writes .trinity/seals/<dir>_<Module>.json. The older files are bare <Module>.json.

t27c seal --save specs/ml/layers/avgpool2d_layer.t27
  -> Seal saved to .trinity/seals/layers_Avgpool2d.json
     .trinity/seals/Avgpool2d.json   untouched, still stale

Re-sealing 94 specs moved the stale count from 136 to 131.

The census

count
stale seals 121
…with a prefixed twin 99
…twin names the same spec_path 97
…twin is current (its spec_hash matches the file today) 81
…twin is itself stale 18
no twin at all 22

The 81 are superseded records whose replacement is present and correct.

Two that are not twins

Bridge_Testbench.json name-matches testbench_APB_Bridge_Testbench.json, which names a different spec_path. A bulk rule keyed on the filename would have treated two different records as one. Any cleanup has to compare spec_path, not names.

Why this is filed and not done

Deleting a seal is a distinct failure class in this gate's own doctrine — DEPARTED, described as "a reproducibility record deleted rather than fixed". Removing 81 at once, plus the ledger lines that reference them, is a coordinated change to the project's reproducibility record and an owner's decision rather than maintenance.

Options:

  1. Delete the 81 superseded seals and their ledger lines in one commit. Gate goes to 40 stale.
  2. Re-seal the 18 stale twins first, then delete 99.
  3. Leave them and baseline the lot — which makes the gate green while 121 records stay wrong, and is the option this campaign would argue against.

Also worth deciding

t27c seal --save rewrites sealed_at on every run whether or not any hash changed. 71 of the 94 files it touched here differed only in that timestamp. A re-seal that changes nothing should probably leave the file alone, or the next bulk operation buries its real content in noise.

Refs #2474

Boundary

  • specs/ml/layers/avgpool2d_layer.t27

Activity

  1. gHashTag commented on Aug 23, 2026

    @gHashTag
    CollaboratorAuthor

    Option 2's first half is not available: 14 of the 18 twins do not parse

    Measured. t27c seal --save on all 18 stale twins:

    sealed cleanly:  4   (and all four changed only sealed_at -- no hash moved)
    rejected:       14   "4 of 4 backends rejected it"
    

    Net movement on the stale count: zero.

    All 14 rejected specs do not parse at all:

    specs/tri/collections/{bitset,lru,map,queue,ring_buffer,set,skip_list,stack,variant}.t27
    specs/tri/graph/graph.t27
    specs/tri/pipeline/builder.t27
    specs/github/tests/e2e_full_flow.t27
    specs/portable/relay_observer.t27
    compiler/codegen/verilog/codegen.t27
    

    Sealing requires generating; generating requires parsing. The seal backlog is downstream of the parse backlog, not beside it. These 18 are stale for the same reason the corpus is 131 source specs short — not because anyone forgot to re-seal them.

    What that leaves

    • Option 2 ("re-seal the 18 first, then delete 99") loses its first half. Those 14 cannot be sealed until they parse, and t27c backlog reports 0 specs at depth 1 — "No single compiler fix can raise the compiling count" — so this is not a near-term unblock.
    • The 81 superseded seals with a current twin naming the same spec_path remain a real and separate bookkeeping question, unaffected by any of the above. That is still the decision this issue is waiting on.
    • coverage will stay red regardless of what is decided here, because the parse backlog sits under it.

    Refs #2479

  2. added
    needs-boundaryNo '## Boundary' section, so the Queen can reserve nothing for it and no bee can take it
    on Sep 23, 2026
  3. gHashTag commented on Sep 23, 2026

    @gHashTag
    CollaboratorAuthor

    A boundary drafted from the paths this issue already names, so a bee can take it. Nothing is reserved until this is in the issue body - the Queen reads the body, not comments. Check the paths first: a wrong boundary reserves files the work does not own.

    Boundary

    • specs/ml/layers/avgpool2d_layer.t27
  4. added a commit that references this issue on Sep 23, 2026
    7a20546
  5. removed
    needs-boundaryNo '## Boundary' section, so the Queen can reserve nothing for it and no bee can take it
    on Sep 24, 2026
  6. gHashTag commented on Oct 5, 2026

    @gHashTag
    CollaboratorAuthor

    Branch queen-2477 deletes about 530 seal files in bulk and adds five Python and shell cleanup scripts plus a 6930-line analysis JSON at the repository root. The scripts are hand-written foreign code that specs/policy/own_language.t27 forbids, and the bulk deletion would trip the seal gate's DEPARTED check for every baselined seal it removes. The project has since settled on keeping twin seals identical with tri seals sync-twins, so the branch should be dropped rather than landed.

  7. gHashTag commented on Oct 5, 2026

    @gHashTag
    CollaboratorAuthor

    Closing as obsolete. The owner decision this issue asked for was taken a different way: twins are kept and kept in step. #2767 added tri seals sync-twins, and the seal-coverage gate names it as the repair ('t27c seal --save && tri seals sync-twins'). On master 13e212a the twin seals are byte-identical to their prefixed siblings in spec_hash and all four gen_hashes (556 such copies), so none of them is a stale superseded record any more. The 39 seals that do not hold are carried in tools/seal_baseline.txt with their causes. If the duplicate files themselves should go, that is a new owner decision about the ledger and naming scheme, not this census. The note that t27c seal --save rewrites sealed_at on every run is a compiler (Rust) change and needs its own owner-approved card.

    Review-jam triage 2026-10-05 (lane B): this card sat in the Queen's review column holding its file boundary. Reopen if this is wrong.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions