workflow-fix: vectorized rewrites land on main + tombstone serial twin

Overview / Motivation

Auto-filed by the workflow-fix-on-bug protocol from a user-approved transcript audit (2026-07-02 PM-chat session; user: "parallelization/vectorization is not properly being used ... Make all seven changes"). Emitting agent: orchestrator, via 5 parallel transcript-mining subagents.

Goal

Require every vectorization rewrite to land the batched helper on main in the same round AND tombstone the superseded serial entrypoint, so reusers and same-issue follow-ups cannot silently re-run the serial original.

Workflow gap

• Bug observed: three build-but-strand / duplicate events in one week: (1) #722's vectorized helper was built on the unmerged vectorized-mlp-skill branch (commit 452c539c31) and then merge-FAILED to main (epm:merge-failed ... reason=new-shared-src-infra-cannot-land-via-artifact) while the #722 session kept running the OLD serial script (~38h ETA); (2) #778's same-issue follow-up RE-RAN the serial 1000-draw null battery on 2026-07-02 because the batched perm_null_draws was not yet on main (confirmed in #836: "the batched perm_null_draws isn't on main, so the rule flags the serial original") — user caught it twice; (3) #834 and the #778 fix session independently vectorized the SAME module in parallel, discovering the overlap only at ~10:09 ("a parallel fix #834 independently landed the equivalent on main").
• Why it is a workflow gap: .claude/rules/workflow-fix-on-bug.md § "Built-but-stranded fixes don't help" states the lesson but prescribes no mechanism; vectorize-many-cell-fits.md names the canonical helper but nothing prevents a still-importable serial twin from being called by the next reuser, and nothing makes an in-session emergency fixer check for an already-open fix task on the same file.
• Confidence (emitter): high.

Proposed change (candidate diff sketch — refine in planning)

In .claude/rules/vectorize-many-cell-fits.md, add a "Supersede contract" section:

+ ## Supersede contract — land on main + tombstone the serial twin
+ A vectorization rewrite that supersedes a serial helper MUST, in the SAME
+ round: (1) land the batched helper on main (not a worktree artifact — the
+ #722 merge-fail is the incident); (2) TOMBSTONE the serial entrypoint:
+ either delete it, or make it emit a loud DeprecationWarning (or raise on
+ EPM_FORBID_SERIAL_FITS=1) naming the batched replacement, so a reuser /
+ same-issue follow-up cannot silently re-run it (#778's follow-up re-ran
+ the serial null battery; #667 called #722's old fit_cell).
+ (3) BEFORE starting an in-session vectorization fix, check for an open fix
+ task on the same file (task_workflow.is_open_workflow_fix_task + a grep of
+ proposed/running infra titles) and adopt/coordinate instead of duplicating
+ (#834 vs the #778 fix session duplicated the same rewrite).

Cross-reference the contract from workflow-fix-on-bug.md § Built-but-stranded (one pointer line).

Scope / surfaces

• Primary targets: .claude/rules/vectorize-many-cell-fits.md, .claude/rules/workflow-fix-on-bug.md
• NOTE: open task #869 also edits vectorize-many-cell-fits.md (widens the trigger to dense linear-algebra fits). Distinct concern, distinct fingerprint — coordinate merges via the normal worktree rebase; if #869 lands first, rebase over it.

Constraints / invariants

• Workflow-surface only. Actual tombstoning of specific serial helpers (e.g. issue722_skill_over_mean.py) is follow-on implementation the rule mandates for FUTURE rewrites; this task may tombstone at most the two known live offenders if trivially safe, else file them.
• scripts/workflow_lint.py default run passes; keep .claude/rules/LESSONS.md index row wording in sync.
• Recursion guard: this session must not auto-route its own subagents' workflow-fix candidates.

Provenance

• workflow_fix_target: .claude/rules/vectorize-many-cell-fits.md, .claude/rules/workflow-fix-on-bug.md
• fingerprint: 75b60cab8677

target_file: .claude/rules/vectorize-many-cell-fits.md, .claude/rules/workflow-fix-on-bug.md bug_observed: #722's vectorized helper merge-failed and stranded off main; #778's same-issue follow-up re-ran the serial null battery because the batched fix was not on main; #834 and the #778 fix session independently duplicated the same vectorization why_workflow_gap: the built-but-stranded lesson is documented but has no mechanism — nothing forces the batched helper onto main, nothing disables the serial twin, and nothing makes fixers check for an open fix task before duplicating work proposed_change: require a vectorization rewrite to land the batched helper on main in the same round and tombstone the superseded serial entrypoint with a deprecation warn/raise pointing at the batched path, and require fixers to check for an open fix task on the same file before starting diff_sketch: |

• Supersede contract: (1) batched helper on main same round;

• (2) serial entrypoint deleted or DeprecationWarning/raise naming the

• batched replacement; (3) fixers check is_open_workflow_fix_task +
  

• open infra titles before starting an in-session rewrite
  

confidence: high related_task: n/a (user-chat transcript audit, 2026-07-02)