release-watch: synth 0.64.0 fixes a silent WRONG-ANSWER bug on every --relocatable compile — pin bumped, proven behaviour-neutral (AFD-114)

[avrabe]

release-watch: synth 0.64.0 fixes a SILENT WRONG-ANSWER bug on every --relocatable compile — pin bumped 0.60.0 -> 0.64.0, proven behaviour-neutral (AFD-114)

synth#1189 / RQ-64-JOINALIAS. On the ARM direct selector — which synth's own PR
calls "every --relocatable compile per #197", i.e. every lowering jess does — an
if/else join register could ALIAS a live register-homed local. local.get of a
register-homed local pushes the home register uncopied; when that is the then-arm
result of a value-carrying if/else, the join mov R_then, R_else on the else path
WRITES THE LOCAL. Exit 0, no decline, wrong answer.

Reproduced on jess's exact invocation (-t cortex-m3 --cortex-m --relocatable):

0.60.0 (jess's pin) 0.64.0
c: mov r0, r2 6: mov r2, r0 <- home copied to a temp
e: adds r3, r0, r0 e: mov r2, r3 <- join writes the TEMP
^ join wrote r0, which 10: adds r4, r2, r0 <- r0 intact
IS local 0's home

ir0(9,0) returns 10 under the pin where 14 is correct.

WAS JESS EXPOSED? No — and the measurement is not vacuous. Every artifact jess
lowers is byte-identical across 0.60.0 -> 0.64.0: the fused cascade on m7dp
(5362d831..., the AFD-106 baseline) and m4f (90e8fa9b...), gale-nano 0.7.0 on m3,
the f100-init and gust:hal probes, and the self-contained passthrough — which also
still matches its COMMITTED .elf (13495328...), i.e. the image that has actually
run on F100 silicon is unchanged. POSITIVE CONTROL: a module carrying the #1189
shape DOES differ (504 -> 510 B). synth's fix changes output only where the bug
fired, so byte-identity across every artifact means the shape does not occur in
anything jess lowers.

THE PIN IS BUMPED, REVERSING THE STANDING POSITION. AFD-106 and AFD-108 declined
bumps because output was byte-identical and a bump widened the varve-layer gap.
Those were no-regression watches. This one is different: the pin CONTAINS a silent
wrong-answer class on jess's exact path, and the bump is provably behaviour-neutral
today — verified end-to-end through tools/deps/fetch.sh + check.sh and a full
cascade rebuild (exit 0, 89,816 B linked, 0 undefined, 5/5 key symbols, fusion
still byte-reproducible). Zero delta now, class excluded later.

Two pieces of accretion removed in the same change:

• the pin directory was named fg60, encoding 0.60.0 in a path — the
  version-in-a-filename hazard AFD-106 flagged and did not fix. Renamed to
  synthpin across all 16 references in 6 files.
• CI fetched synth 0.62.0 SEPARATELY, only because the old pin predated
  verify-embedder. The pin now carries it, so that download was a stale version
  left behind by an earlier bump. Removed; one synth builds and checks. Verified
  the pinned binary accepts with --allow-writer (0) and refuses without it (1).

WHAT 0.64.0 DOES NOT FIX: synth#1191 (filed by jess at 14:54, before this 18:42
release) is untouched — gale-nano's lowering is byte-identical to 0.60.0, same six
undefined symbols, still no symbol for the gust:os/time export the component
declares. H3's second half remains blocked there (AFD-113).

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