fix(dev/ds compiled-runner): pipe-safe parsing, N/A-test convention sync, ds outputsProduced forcing function (v5.68.3)

[edwinhu]

Summary

Fixes 8 verified code-review findings on the pass-#9 shared compiled-runner core (scripts/lib/plan_table_core.py, workflows/templates/run-core.js, and the dev/ds fragments/compilers). Version bump assumes PR #50 (5.68.1) and the pending round-2 PR (5.68.2) land before this one — rebase/renumber if that ordering changes.

What / why, per finding

1. Pipe-safe cell splitting (high reachability). plan_table_core.split_row/find_table split on bare |, so a Verify cell with a shell pipe (`pytest -q | tail`) or regex alternation (`grep -E 'foo|bar'`) shifted every later column — this was actively corrupting the repo's own .planning/PLAN.md. New _split_cells() helper is aware of escaped \| (unescaped to a literal pipe) and pipes inside backtick code spans; both split sites now share it.

2. dev N/A-test convention drift. dev-task.js's testRequired() and the parser's now-deleted _TEST_NA disagreed on em/en-dash cells (—/-): the parser treated them as N/A, the compiled gate's predicate did not — a dash Failing-Test cell compiled but could never satisfy its own gate. dev-task.js is now the sole owner of the predicate (cross-referencing comment in both files); the parser's dead test_required/_TEST_NA are removed.

3. ds prefix-tolerant header check. ds's required-column check used exact c not in header while cell extraction is prefix-tolerant (dev already used the shared has_col()) — a qualified header like Expected Output (per requirement) was falsely rejected as missing. Now uses plan_table_core.has_col().

4. Restored ds's outputsProduced forcing function. The run-core extraction silently weakened it from schema-required + === true to optional + !== false. Restored via a compiled config hole __REQUIRE_OUTPUTS_PRODUCED__: ds_compile.py sets True, dev_compile.py sets False; run-core.js gates on it in both TRANSFORM_SCHEMA.required and the pass/collect logic — no domain names leak into the shared core.

5. Deps-cell phantom-dep guard. parse_deps swept the whole cell for digit-looking tokens, so free text like T1 (needs config v2) created a phantom dep on v2. Now splits on commas/whitespace (after stripping after /backticks) and requires each piece to fullmatch a dep token; non-conforming residue is dropped and, via an optional violations list threaded from both domain parsers, reported instead of silently becoming a dependency.

6. Mixed-level skip stamping. run-core.js's mixed-level branch (some tasks done, some todo) never stamped a state entry for blind-skipped done tasks, so collect() under-reported tasksDone/over-reported tasksRemaining and scoreTable showed · for completed work. Now stamps the same skip-state shape the all-done branch uses.

7. Cleanup. Dropped the dead cfg.clearedRecheck alias (no producer exists). recheckTrigger().atLevel was always an echo of the li argument — the core now supplies atLevel: li directly; dev-task.js no longer returns the field.

8. Guard-hook cleanup. dev-plan-executable-guard.py's docstring wrongly claimed it's wired via dev-plan-reviewer frontmatter (that skill is read-only, never writes PLAN_REVIEWED.md) — corrected to dev-design/SKILL.md. Both guard hooks now import their parser lazily inside validate_plan(), so the hundreds of unrelated Write/Edit hook firings per session skip the parser import.

Deferred (not attempted)

• Extracting the duplicated compile driver (dev_compile/ds_compile) and guard shells into shared modules
• Dataflow (non-barrier) scheduling in run-core
• The pre-probe skip optimization
• ds_compile._tier keyword-sniffing → a real Tier column
• Unifying the Python/JS toposort tie-breakers — added a code comment at run-core.js's toposort noting it must stay order-compatible with plan_table_core.toposort_ids

Test plan

• uv run python3 tests/plan_table_core_test.py — 30 passed, 0 failed (10 new)
• uv run python3 tests/dev_plan_table_test.py — 38 passed, 0 failed (7 new)
• uv run python3 tests/ds_plan_table_test.py — 17 passed, 0 failed (2 new)
• node tests/dev-run-driver.test.mjs — 35 passed, 0 failed (pre-existing, unmodified — proves no regression)
• node tests/ds-run-driver.test.mjs — 28 passed, 0 failed (pre-existing, unmodified)
• node tests/ds-grain-pause.test.mjs — 17 passed, 0 failed (pre-existing, unmodified)
• py_compile on all 7 edited .py files
• node --check on run-core.js, dev-task.js, ds-task.js
• End-to-end compile smoke: synthetic PLAN.md fixtures (piped verify cell, em-dash test cell, qualified ds header) compiled via both dev_compile.py and ds_compile.py — emitted run.js node-checks clean and contains the correctly-parsed verify strings + the compiled REQUIRE_OUTPUTS_PRODUCED flag (false for dev, true for ds)
• Re-read every edited template-literal in run-core.js/dev-task.js for unbalanced backticks (known incident: node --check doesn't catch this)

🤖 Generated with Claude Code

https://claude.ai/code/session_01J3NEAUUosUHyci4iFZBH2J