fix: the off-pin detector judged a PATH STRING, so the pinned binary was declared a candidate differential (AFD-115)

[avrabe]

fix: the off-pin detector judged a PATH STRING, so the pinned binary was declared a candidate differential (AFD-115)

Clean-room verification of AFD-113/AFD-114. Every substantive claim in both
CONFIRMED — six artifacts byte-identical across the pin bump, the #1189 signature,
the silicon result, gale-nano's import/export asymmetry — and neither finding
overstated scope. What the audit found was in the MACHINERY, again.

1. THE OFF-PIN DETECTOR JUDGED A PATH STRING, CONTRADICTING ITS OWN COMMENT.
   cascade-invoke/build.sh used [ "$SYNTH" = "$SCRATCH/synthpin/synth" ], an
   absolute-path compare. Invoking the GENUINELY PINNED binary by a relative path —
   what a local run types, and what AFD-114's own verification used — took the off-pin
   branch, announced a CANDIDATE differential for an on-pin build, and made
   REQUIRE_ON_PIN=1 FAIL on the correct binary. Six lines above, the meld/loom comment
   already says why: "judge by DIGEST ... otherwise every CI run would announce itself
   as a candidate differential and the banner would stop meaning anything."
   announce_tool does it right; the synth branch bypassed it. CI escaped by accident
   (ci.yml omits SYNTH, so the default absolute path string-matched). Now a digest
   comparison; verified both ways.

2. A VACUOUS AGREEMENT CLAIM I INTRODUCED YESTERDAY. AFD-114 removed ci.yml's
   SYNTH_VERSION as accretion; check-drift greps for that literal, so synth and loom
   now read - in the CI column — while the summary still said "ci.yml and the varve
   pin agree" and "every tool compared (5)". No such comparison happened for those
   two. Now counts only where BOTH are declared and NAMES the uncompared tools.
   Blocking class re-verified potent.

3. THE #1189 EVIDENCE COULD NOT BE RE-RUN. The load-bearing "the control fires"
   half existed only as prose. tools/lowering/joinalias/ commits the module and
   asserts the property directly on the pinned synth. Potent both ways: PASS on
   0.64.0 (adds r4, r2, r0), FAIL on 0.60.0 (adds r3, r0, r0). Wired into CI.

4. THREE FIXES IN THE GALE-NANO HARNESS.
   (a) the seam check asserted membership, not exclusivity. Writing the exact-set
   assertion surfaced a real subtlety: nm lists all SIX as undefined even after
   aliasing, because --redefine-sym renames the DEFINITION to the imported name
   so the object carries both T and U. The raw U list is NOT the obligation.
   Now compares UNRESOLVED = undefined MINUS defined; breaking an alias is
   refused.
   (b) the gale-nano digest was hardcoded — a second copy of a value that also lives
   in artifacts.pins with nothing enforcing agreement, the drifted-mirror shape
   AFD-104 deleted a registry for. Now read from the pins file.
   (c) run-on-silicon.sh printed handle and exec_state but never asserted them, so
   the exec_state==1 that localises the fault to poll-round was eyeballed. Both
   asserted now; re-run on silicon, result unchanged.

Also corrected in AFD-113/114 themselves: "8/8 assertions" (the oracle emits no
such count), citing jess's OWN WIT comment as the record for handle 0 (circular —
the fact is measured, the citation was not evidence), and "16 references" where
the count is 20.

Co-Authored-By: Claude Opus 4.8 noreply@anthropic.com