TASK-121 — a check that reads live project state as its expected value is now found mechanically

[ranjiao]

TASK-113 found five of these by hand in one afternoon and the pass was thrown away. Three more arrived within the week — two the moment .perry/config.md declared its first track, two when PR #14 changed which paths the schema says Perry owns. There was no mechanism; there was a memory of having looked. This is the mechanism.

What I decided the class is

A check is in this class when a value it read out of the project it lives in is asserted equal to a literal that enumerates or counts what that project happens to hold today. Both halves must hold:

1. it reaches live state — a path schema/state-schema.json declares Perry writes, resolved through the State root: line in .perry/config.md, or the parsed payload of a bin/ tool run with cwd=/--root pointing inside this repository. The path list is read out of the schema and never named in the guard: instance 8's literals were about which paths Perry owns, so a guard keyed on BOARD.md or the task store would have missed it. test_the_same_claim_lands_where_the_state_root_points builds two projects from one schema and shows the same claim landing at BOARD.md and at docs/BOARD.md.
2. its expectation is closed — a non-trivial constant, or a non-empty list/set/tuple/dict display, including one reached through a module or class constant or through set(...). [], {}, 0, 1, "" are not closed: "nothing is wrong" is a property quantified over whatever the project holds, and it is the shape every one of TASK-113's repairs converged on.

Half one keeps test_prioritize — which asserts exact rendered tables against boards it wrote itself — off the report. Half two keeps sum(...) == len(records) and problems == [] off it.

What it deliberately does not catch

The module docstring names seven, including two of the eight known instances:

• containment — instance 8's assertIn("perry/tasks.jsonl (unclaimed)", …). Growth cannot falsify a containment assertion, and test_md_store kept exactly that shape through its repair; catching instance 8 needs a path-literal signal that cannot be told from the hundreds of path strings the fixtures legitimately write.
• a live value used as INPUT — instance 2's TASK-038 borrowed off the live board. The defect is in the fixture, not in an expected value.
• also: assertTrue/assertFalse on a live value; a tool's exit code and human text; code/contracts/templates (schema/, SKILL.md, bin/ — a test SHOULD go red when the schema is edited); the repository root handed to a helper (tried, reverted — it put 21 fixture-only assertions on the report); anything outside tests/.

Verification

1. It finds instances it was not shown. Three are checked in verbatim as whole modules under tests/fixtures/live-state/, each the module as its repair found it:

fixture	commit	repaired by	flagged
md_store.before.py	d90612a (f3c4461^)	f3c4461	report["kinds"] == {"setting": len(records)}
track_attribution.before.py	d90612a	f3c4461	[t["track"] …] == ["main"]
v5_signoff.before.py	e116f8a (cbbc41a^)	cbbc41a	set(events) == set(self.SIGNED)

Whole modules, not excerpts — trimming is hand-writing an approximation by another name, and the analysis reads module bindings, class attributes and sibling methods. test_the_fixtures_are_what_git_holds byte-compares against git show where the history is reachable; test_the_fixtures_have_not_been_edited pins SHA-256 for CI, which checks out at depth 1 and cannot.

2. It does not flag the fixes. tests/test_track_attribution.py and tests/test_v5_signoff.py as they stand today are clean end to end — and the repaired TestAProjectWithNoRegisterIsUnmoved still asserts ["main"], against a project it built itself. test_md_store's repaired assertion is gone from the report.

3. False-positive floor: 6, stated, over 65 modules and 1930 tests. Recorded with a verdict apiece in tests/fixtures/live-state-expectations.json — a new hit is a red, not a line in a report nobody reads.

Three real instances (each owes its own row; this row fixes none of them):

• test_md_store § test_okr — assertGreater(len(krs), 20) over perry/OKR.md. c9018ae's shape exactly.
• test_task_writer § test_every_hand_written_row_in_perrys_own_board_round_trips — assertGreater(len(rows), 5) over Perry's own BOARD.md.
• test_prioritize § test_an_id_shaped_word_in_prose_is_warned_about — ctx built from load_task_records(perry/). Its neighbour is the more fragile half and is not flagged, because [] is not closed: fn("see ADR-006 and USER-014", ctx) == [] needs both ids to still resolve on this board.

Three false positives, named rather than silenced:

• test_contract_invariance:208 — semantics is a changelog bin/perry-task carries in its own source; half one cannot tell a tool-built payload subtree from a project-built one.
• test_md_store:300 — "2099-01-01" is the value the test wrote into the store four lines earlier.
• test_md_store:314 — state_root is the record the test selected and mutated.

4. Reverting the guard reddens its own test. Five mutations, each a red: dict displays no longer closed (instance 6), no constant folding through a class attribute (instance 1), the state root not resolved (instances 6 and 7), and each half removed in turn.

5. python3 tests/parallel -j 4: 66 modules · 1951 tests · 1 module red — test_diagnose, user_load.dangling non-empty. That is TASK-126 on unmerged PR #22 and it is inherited: my own baseline on 5dd745f before touching anything was 65 modules · 1930 tests · the same 1 red. python3 bin/perry-lint 0 errors (3 pre-existing NS-01 warnings), git diff --check clean.

git diff -- perry/ is empty.

Note on the file count

feat/work-modes is ahead of origin, so every PR cut from it carries those commits. My own code diff is 6 new files, 0 modified — tests/live_state_expectations.py, tests/test_live_state_expectations.py, tests/fixtures/live-state-expectations.json, and three history fixtures.

🤖 Generated with Claude Code