[dev-pmallapp]

Closes #325. Refs #293. The last task of S2.6.

Build target 10. Cases 3 and 19.

Line-neutral, measured properly

File	Before	After
docs/milestones/M2.md	261	261
docs/milestones/M4.md	259	259

Proven with awk 'END{print NR}' against git show HEAD:<file>, not wc -l alone — wc counts newlines, so a file whose final newline changed can read identically while its line count did not. 7 insertions, 7 deletions, four hunks.

The four restatements

M2.md:107-109 — "the second run makes zero mutations" with no > 0 beside it is satisfied perfectly by a harness that never ran anything: the exact vacuity #322's runner was built to close, left standing in the sentence that seeds it. Now per scenario and never summed — run 1 makes at least one mutation, run 2 then makes none.

M2.md:105 — iai-skills is a docs target with no build file. Gate G5 made the harness its own target; the milestone now names iai-harness, checked against the row #318 added at ARCHITECTURE.md:62 rather than assumed.

M2.md:230-232 — the bare-token ban was overturned by gate G2 of #41 as unimplementable and self-contradictory. Now bans the routing form. Second time the file was opened with this superseded; fixed, not deferred a third time.

M4.md:215 — records the one fixture repository S2.6 owns and M4 consumes, per gate G6.

The point of the task: eight citations now resolve and are false

#297's edit was exactly line-neutral so that its citations would keep resolving. They do. They are also no longer true. A guard checking only that a :NN resolves would have passed it, because making it pass was the design goal.

Every citation into the four edited sites was resolved by hand. All still resolve. Eight are now false in the present tense — six saying :105 assigns the harness to a docs target or that :230-232 still carries the bare-token rule, one quoting the replaced :107-109 sentence verbatim, one saying the M4.md:215 restatement is unresolved. Each is tabled in the evidence with a verdict, alongside the six that survive and why.

None is edited. Every one is a problem statement or gate ruling inside gate-approved documents that record the world as it was at ruling time. A Design that stopped describing the defect it commissioned a fix for would destroy the record of why the fix happened.

No immutable evidence artifact cites any edited site — checked across all of docs/evidence/.

Two things deliberately not done

M2.md:240-241 is untouched, owner #337. #324's token figure is a session total and an upper bound; whether to tick, restate or re-measure is a ruling, not a line edit — and making it silently here would be the third superseded criterion rewritten without one. It must not be ticked on that number.

#304 is not closed. Nothing verifies a :NN still says what the citing document claims. Eight were checked by hand because there is no other way — the third consecutive task to pay that cost.

A drafting error worth recording

The first version of the :107-109 rewrite compressed the tail and silently dropped "every claim maps to at least one case" — a verification requirement, not a turn of phrase. Caught by diffing meaning rather than counting lines. A line-neutral edit can be line-neutral and still delete a requirement.

Chain, on a3d02b0

1156 tests / 1142 pass / 0 fail / 14 skip / 4220 expect() · lint 0 across 126 files · typecheck clean · claim-lint 126 files, 0 errors, 1 warning · skill-harness exit 0 · both verifiers PASS at six.

Evidence: docs/evidence/325-20260915T145546Z.md.