Closes #438.

mechanisms.md had no entry for "two stages need the same answer about the same input", though the codebase has executed the pattern five times and decisions.md states the principle twice — each time inside a dated entry about a specific fix, where nobody looks it up.

The entry

ONE-PREDICATE-PER-QUESTION — the stage that does not decide calls the one that does.

The site that does not own the decision calls the deciding stage's own predicate, never a condition written to match it. A hand-written mirror agrees with its original until one of them moves, and the drift is invisible in both directions: each site keeps passing its own tests while they disagree about an input neither covers.

The five instances, each found as a defect before it was found as a pattern:

The Known-limit clause is the half worth writing down

The share is only available in the import direction. _assign imports _group, so a piece-level predicate both stages ask has to live in _group — not because grouping owns it. A text-level one goes to _vocab instead: the destination follows the layer, not the topic.

Where the reader comes after the decider, the alternative is to record the answer on the state — ParseState.order is exactly that shape, and says so in its own docstring. That option is unavailable whenever the earlier stage is the one asking, which is what made #429's design a discovery rather than a lookup.

One citation added

_group.py's module docstring restated the import-direction claim in free prose — which the mechanisms preamble names as the failure mode: "Free-prose restatements of mechanism claims are how one wrong sentence has shipped six times." It cites the slug now.

Two claims corrected while writing it

Both were carried from session notes rather than the committed record, and both were caught by running axis 1 (recompute every number) and axis 4 (claims about a document vs what it says) on my own draft:

• reads_as_suffix no longer exists. It was real in fix(group): the bound-given reserve and join ask what assign reads (#401, #421) #423, but abdul Smith Jr Ma loses the family name — P5's reserve does not mirror S2's bare-acronym fork #425 superseded it with the shared peel. That pair's lasting artifact is is_trailing_numeral_suffix, which is what the entry names. (The originating issue body has the stale name.)
• "four inexact-mirror drafts" and "twice the right answer" are counts no committed document supports. The record carries the lesson — mirror assign's walk, not its condition — but not the tally, so the entry states the lesson without a number.

Verification

Docs-and-one-comment only. Suite 5668 passed, mypy and ruff clean, differential 0 unexplained at the default baseline (unchanged, as expected for a change that alters no parse).

Axes run: 1 (recompute — found both corrections above), 2 (no detector used), 4 (every quoted excerpt checked verbatim against its source — #319's decisions line, ParseState.order's docstring, the mechanisms preamble), 7 (the "Reach for it when" clause is advisory and sweeps in nothing), 9 (the new anchor resolves; the citation test passes). Axes 3, 5, 6 are rules.md-specific and inapplicable — no rule statement, example or interacts: line changed. Axis 8 inapplicable: no guard added. Axis 10: the entry's only capability claim is that _assign imports _group and cannot be imported back, verified.

🤖 Generated with Claude Code