feature: gale-nano 0.7.0 EXECUTES on real STM32F100 silicon — admit works, dispatch does not (AFD-113)

[avrabe]

feature: gale-nano 0.7.0 EXECUTES on real STM32F100 silicon — admit works, dispatch does not (AFD-113)

H3 second half: ADVANCED, NOT COMPLETE. Stated that way deliberately.

gale-nano 0.7.0 (digest-pinned) lowered for cortex-m3 and linked against jess's
real gust:hal MMIO (AFD-109), the generic embedder-init apply loop (AFD-107) and a
boot shim. 4,349 B, 0 undefined, embedder ABI verified. On silicon:

LEG (only the ADMIT word differs) poll-task# handle state completion
task admitted (baseline) 00000000 00000000 00000001 c0ffee00
no admit (negative control) 00000000 deadbeef deadbeef c0ffee00

Gale's code RUNS ON THE PART and exec_admit WORKS: gale's own exec_state(handle)
returns 1, exactly what TEST-PIX-035 asserts. handle 0 is correct, not a failure —
jess already recorded that gale-nano hands out handle 0.

But exec_poll_round reaches poll-task ZERO times, where the wasmtime reference —
re-run green the same day, 8/8, same pinned artifact, same PRIO/DEADLINE/NOW
constants taken from app/dispatch-driver — reaches it once. That is a confirmed
wasmtime-vs-silicon divergence, which is what jess's lane exists to produce.

LIKELY CAUSE, NOT PROVEN, AND NOT A GALE DEFECT. gale-nano's component manifest
declares exactly two imports (gust:os/taskdisp, gust:hal/mmio) and exports
gust:os/{time,log,spawn,exec,timer} — matching gale#223 in words. But
synth --relocatable --all-exports emits SIX undefined symbols and an export set
that does not match the declared one: it emits gust:sched/tasks@0.1.0#*, which the
component does NOT export, and nothing for gust:os/time@0.1.0, which it DOES.

Nothing in the object separates the 2 real obligations from the 4 internals, and
satisfying them the obvious way is measurably wrong:

• stubbing set-deadline/slept-status/state inert replaced gale's task-state
  store. Aliasing them onto gale's own exports via objcopy --redefine-sym is
  correct and is what build.sh does.
• deadline has nothing to alias to, because the gust:os/time export was not
  lowered. Returning 0 broke dispatch; a now+ticks guess did not fix it either.

jess is not going to keep guessing at a gale semantic — that is gale's lane, and
inventing one is how a stub becomes a false claim.

Two harness defects of jess's own, both caught by the reference rather than by
inspection:

• the first run used deadline=(0,0), now=(0,0); the reference's own comment says
  NOW must be PAST the deadline for a task to be READY. Reporting that as
  "gale-nano does not dispatch on silicon" would have been a false supplier
  report — the meld#390 shape.
• the openocd readback filter ^0x2000048 silently dropped the exec_state word
  at 0x20000490 — the one reading that localises the fault.

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