Skip to content

seals: re-seal ml_train_hslm.json -- spec changed since sealing, coverage red on PRs #13957

Closed

Description

@gHashTag

Boundary

  • .trinity/seals/ml_train_hslm.json

User Scenarios

  • Given master merged spec changes to specs/ml/train_hslm.t27 without re-sealing, when any PR runs the coverage check, then it reports ml_train_hslm.json [stale] ("specs/ml/train_hslm.t27 changed since sealing") and fails.
  • Given the seal is re-saved from a master-built compiler on a branch containing master, when the coverage check runs, then ml_train_hslm.json holds and the check passes.

Requirements

  • FR-001: .trinity/seals/ml_train_hslm.json MUST be re-saved with a master-built t27c from a tree containing current master (t27c seal specs/ml/train_hslm.t27 --save).
  • FR-002: No other seal, spec, or file MUST be touched.
  • FR-003: The seal MUST NOT be baselined (seal_baseline.txt) -- baselining records drift as accepted debt, which the coverage tool itself names as the wrong repair.

Success Criteria

  • python3 tools/check_seal_coverage.py reports ml_train_hslm.json as holding (exit 0 on that seal).
  • coverage check passes on the fix PR.
  • git diff --name-only origin/master lists exactly .trinity/seals/ml_train_hslm.json.

Activity

  1. gHashTag commented on Nov 15, 2026

    @gHashTag
    CollaboratorAuthor

    Finding (deeper root cause, 2026-11-15)

    Re-seal attempt with a master-built t27c (release, commit 6e0e2c4) was
    refused by the compiler itself:

    refusing to seal specs/ml/train_hslm.t27: its seal records a pass and it no
    longer compiles (does not compile: spec.zig:1761:40: error: type 'f64' does
    not support array initialization syntax)
    

    The spec's line 1931 (var win_g : [BATCH * NPARAM]f64 = [0.0] ** (BATCH * NPARAM);,
    1.1 GB zero-init, deliberate per its comment) hits the known gen-zig
    f64-array-repeat gap -- the same family as #13683 ("module-level array buffer
    ([0] ** N) while gen succeeds") and the @splat migration (#13591/#13625).

    The seal cannot lawfully be re-saved until the spec compiles again; the fix
    belongs in the compiler (gen-zig), not in the spec. This issue stays open as
    the seal-drift tracker for ml_train_hslm.json; the compiler gap should be
    resolved under #13683's family, after which a re-seal (this Boundary) closes
    this issue.

    PR #13953 (sieve_design_law + glass_window) covers the other two drifted
    seals; ml_train_hslm was the uncovered third.

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