TASK-141 — a stored blocked no longer masks an empty blocked_by

[ranjiao]

The bug

startable read the row's own status before the dependency graph it had already computed one loop above:

bin/perry-task:4728
task["startable"] = bool(task["open"]
                         and task["status"] not in {"blocked", "review"}   # read first
                         and not task["blocked_by"])                        # never reached

so the third term was unreachable for exactly the rows it existed to catch. On this project's own board TASK-037 (waiting on the closed TASK-092) and TASK-045 (on the closed TASK-044 → TASK-047 chain) both reported

status=blocked   blocked_by=[]   startable=False

in one object — two of the four blocked rows, one still carrying a Next action naming a chain that had fully closed. The other half: perry-task done never looks at its dependents, so the ordinary close path creates that state and this line then hid it.

The option taken, and what the other would have cost

Option 2. The read path stops deferring to a status the graph contradicts, and the payload names the disagreement instead of silently recomputing it. New tasks[].blocked_stale is true when a row is open, stored blocked, declares at least one dependency, and every one of them has closed.

The stored status is deliberately left alone. list reports; it does not rewrite a cell nobody asked it to rewrite. So a stale row still reads blocked on the Board and in status until somebody acts — blocked_stale is what makes that visible meanwhile. That is the cost, stated rather than hidden.

What option 1 (the close path clears its dependents) would have cost. It writes state the user did not ask for, on rows the user did not name, as a side effect of closing a different row — and to be honest about it every such write needs its own event in the log (V3 item 4; TASK-139 is open for exactly that shape), so it is a write-path change with a log-format obligation attached, not one computation. It also fixes nothing already on disk: every board that is currently stale stays stale until each row is closed again, and any store written by hand, imported, or migrated re-creates the condition. Option 2 corrects every such row on the next read, on boards that already exist. The two are complementary rather than alternatives — if option 1 ever lands, blocked_stale is what would let you verify it worked.

Not "delete the check"

Three cases unchanged, each tested in both directions:

• a row with at least one open dependency keeps a non-empty blocked_by and stays unstartable — TASK-050 (on TASK-094) and TASK-067 (on TASK-094 + TASK-095) are the live instances, and both still report startable=False blocked_stale=False;
• a blocked row that declares no dependency at all is untouched — its blocker is prose Perry cannot read (conformance.blocked_without_dependency), and "I cannot see it" is not "it closed";
• review is untouched — a row waiting on a human can never be contradicted by a dependency edge.

Verification

Every fixture is built through the store's own writer (add, status --status blocked --on, done) against a temp root — never by hand-editing a board, which would be drift and would not exercise the close path that produces the bug.

tests/test_stale_blocked.py, 11 tests. Proved by mutation in both directions:

• reverting startable to the pre-TASK-141 formula reddens exactly 2 tests, both the new case (test_the_measurement_itself_no_longer_reproduces, test_closing_the_second_blocker_frees_its_dependent_too) and leaves the genuinely-blocked and boundary tests green;
• deleting the status check outright — the wrong fix — reddens exactly the two boundary tests (test_a_blocked_row_that_declares_no_dependency_stays_blocked, test_a_row_in_review_is_untouched).

Contract

Moves to perry-task/list/1.12 with a semantics entry, because a changed meaning is not covered by "1.x only adds keys". The key-parity metric is unchanged at 0 documented-not-emitted / 17 emitted-not-documented — the new key was documented in the same edit; the baseline diff is the version string plus 87→88 documented / 99→100 emitted.

blocked_stale is derived on every read and never reaches tasks.jsonl — perry_store.record() uses a fixed STORED allowlist, and a test pins that.

Out of scope, respected

conformance.blocked_by_closed_rows is TASK-142 and is not added here. git diff -- perry/ is empty; TASK-050 and TASK-067 were not touched.

🤖 Generated with Claude Code