The pin says which voice answered

[felipesauer]

What

The guard that walks the surface asking which paths stamp a run has one way of being wrong it cannot see: a synthesised line the declaration could not answer for is refused before any action runs, and it then reads as a path that stamps nothing. The previous slice wrote a pin for exactly that, and the pin looked for the literal Usage: mnema.

Half of the premise under it was false, and it was the half that matters. The misuse voice does carry the usage line the help prints — the same text, usageOf(command) — but Usage: is added outside it, by commander's help formatter. The misuse voice indents the line with no prefix at all. Measured against the built binary:

$ mnema task
mnema task needs <title>: the task title
  mnema task [options] [command] <title>

So the pin saw the help and none of the seven codes the surface words. mnema task, mnema decision <title>, mnema link a b --rel and mnema completion powershell each came back asked: false and parserAnswered: false at once — precisely the silent misclassification the pin exists to catch.

How

• The discriminant calls the function the writing calls. usageOf is exported with the reason on it, and theParserAnswered lives in the walk's own support module — one site, read by both files that ask.
• It covers the parser's third answer. A code nobody worded carries no usage line at all and takes the same shape as a verb's own refusal (Refused (…): …), so it is told apart by the code's namespace, which is an identifier and not prose.
• Both measuring sites build their row through one function. They were four copied fields, which is the shape that let the classification be wrong in two places at once without either being wrong on its own.
• Presence is asserted before absence. The case requires the usage line to be on the stream, and only then requires Usage: not to be. An absence alone goes vacuous the day the invocation stops printing.

The rule had a second site

outside-a-project-the-surface-says-so.test.ts classifies the same three paths as answered by the parser, through a hand-kept table whose reason was three copies of a sentence nothing read — a reconciliation of names can say a path refused something else first without ever reading what it said. That reason is now the one entry in the table a guard can check, asserted in both directions: a path declared as answered by the parser has to have been, and no path that missed the project may look like one.

Reachability of the third arm was measured rather than assumed: commander declares 13 codes, seven are worded, three mean "already answered", and the remaining three (conflictingOption, error, executeSubCommandAsync) are unreachable today — .conflicts(, .error( and executableFile return zero across packages/*/src.

Mutation battery

Thirteen rows on one commit, in the foreground, reds read off the product's own ledger and each re-run alone before it counts. Every reading collected 4413 cases; caught equals in the run on all of them.

#	mutation	expected	verdict	caught
m1	the defect put back: the literal Usage: mnema again	red	RED	1
m2 / m3	each arm of the discriminant lost, one at a time	red	RED	3 / 1
m4 / m5	it says yes to everything / no to everything	red	RED	3 / 3
m6	the one row-builder reads the wrong stream	red	RED	4
m7 / m8	the second site, each half	red	RED	1 / 1
m9	the writing drops the line to type — proves the presence guard	red	RED	8
x1	DIAGNOSTIC: the writing's usageOf gains the help's prefix	red	RED	5
c1	CONTROL, has to light	red	RED	5
c2	CONTROL, has to fail to apply	not-applied	NOT APPLIED	—
c3	CONTROL, applies and has to leave zero	zero	ZERO	0

m1 is the row that decides the change: before this, nothing in the repository lit with that line applied. x1 is the argument for reading the product's function instead of copying its text — a reading holding a copy would light zero there.

Scope

src/ carries one change and it is visibility: usageOf went from private to exported. No verb's behaviour changed. CLAUSES_IN_THE_TREE was restated from 2421 to 2422, as its own doc instructs, because this change adds one import to the test tree.

Validation

pnpm build && pnpm lint && pnpm test over the committed tree, clean status: 281 files, 4413 cases, exit 0.