fix: take the signal-time autosave on the game goroutine (closes #24)

The SIGHUP/SIGTERM handler gob-encoded the live game tree from the signal
goroutine while the game goroutine was mid-turn mutating it, and AutoSave
removed the save file before encoding — so the failure mode was not a
stale save but a deleted one followed by a possibly torn replacement,
with a window in which the player had neither. The suite has run under
-race since 2026-08-09 and was green because nothing had ever driven the
turn loop concurrently with a signal: evidence of untested, not of safe.

The handler no longer writes anything. AutoSaveOnSignal posts a request,
wakes the input read, and waits up to signalSaveTimeout for the game
goroutine to take it; the encode runs on the goroutine that owns the
state, at the three points where that goroutine can sit: between turns
(command), on waking from a blocked readchar, and while parked in the `!`
shell escape (runShellEscape, which now runs the shell on a helper
goroutine so a hangup during it still rescues the game).

Blocked on input is the case that matters — a dropped connection lands
while the player is thinking, so a flag checked only between turns would
never be looked at. Terminal.ReadChar therefore returns (byte, bool),
with ok false meaning "woken by Interrupt, no key", and term.Tcell posts
a tcell.EventInterrupt onto tcell's own event queue to unpark PollEvent.
readchar services the request and reads again, so no caller sees it.

saveFile writes a temporary file in the save's own directory, fsyncs it
and renames it over the target instead of truncating in place, so a save
that fails — or never happens because the deadline ran out — leaves the
player's previous save whole.

The SIGINT/SIGQUIT no-save decision and the single-signal-read ordering
guarantee are untouched. pendingSaver reads the game out from under its
mutex rather than delegating with it held, because the delegated call now
blocks until the save is taken.

TODO.md

@@ -34,6 +34,55 @@ wizard commands).
 
 # Completed Steps
 
+- 2026-08-09 Signal-time autosave moved onto the game goroutine
+  (`fix/autosave-race`, closes #24): the SIGHUP/SIGTERM handler gob-encoded the
+  live game tree from the signal goroutine while the game goroutine was mid-turn
+  mutating it, and `AutoSave` **removed** the save file before encoding — so the
+  failure mode was not a stale save but a deleted one followed by a possibly
+  torn replacement, with a window in which the player had neither. `make test`
+  has run with `-race` since 2026-08-09 and was green, because no test had ever
+  driven the turn loop concurrently with a signal: evidence of untested, not of
+  safe. The handler now writes nothing itself. `AutoSaveOnSignal` posts a
+  request on a one-deep channel, wakes the input read, and waits up to
+  `signalSaveTimeout` (3s) for the game goroutine to take it; the encode happens
+  on the goroutine that owns the state. **The blocked-on-input case is the whole
+  point** — a dropped connection lands while the player is thinking, so a flag
+  checked only between turns would never be looked at — and it is handled by
+  making the read interruptible: `Terminal.ReadChar` returns `(byte, bool)` with
+  `ok == false` meaning "woken by `Interrupt`, no key", `term.Tcell.Interrupt`
+  posts a `tcell.EventInterrupt` onto tcell's own event queue to unpark
+  `PollEvent`, and `readchar` services the request and reads again, so no caller
+  sees the wake-up. The other unbounded park is the `!` shell escape, where a
+  hangup used to save and would otherwise have regressed to not saving: the
+  shell now runs on a helper goroutine and `runShellEscape` selects on {shell
+  finished, save request}, keeping the encode on the game goroutine while it
+  draws nothing. Between turns (`command`) covers a game that is busy rather
+  than parked. The wait is bounded so that a game goroutine wedged with no
+  service point can never stop a signal from getting the process out; giving up
+  costs nothing now that `saveFile` writes a temporary file in the save's own
+  directory, fsyncs it, and renames it over the target instead of truncating in
+  place — a failed or skipped save leaves the previous save whole. New
+  `game/autosave_test.go` drives the real turn loop while a second goroutine
+  asks for 25 saves (the interleaving that never existed before), plus the
+  parked-on-input case with a terminal fake that genuinely blocks, the shell
+  case, the deadline case (previous save byte-for-byte intact), the no-file-name
+  case, and the rename discipline — the last pinned by a handle opened before
+  the save, which still reads the old file whole after it. Each was
+  mutation-proved: reverting `AutoSaveOnSignal` to encode on the calling
+  goroutine (the pre-fix behavior) makes the turn-loop test fail under `-race`
+  with 113 reports, and removing each of the three service points fails exactly
+  the test for that park with its own message. `pendingSaver` now reads the game
+  out from under its mutex instead of delegating with it held, because the
+  delegated call blocks until the save is taken — the PR #23 review's N3 note,
+  load-bearing rather than hypothetical, and pinned by a test. The
+  SIGINT/SIGQUIT no-save decision and the single-signal-read ordering guarantee
+  are untouched; `savesOnSignal`'s third ground ("safety") is rewritten, since
+  the corruption window it weighed no longer exists. `MEMORY.md` stops listing
+  signal-time autosave among the deliberate `_ =` discards and states the new
+  discipline; `ARCHITECTURE.md` §5.3, the `Terminal` sketch, the C-to-Go mapping
+  row and §9's SIGTSTP paragraph are corrected to match. `Next Step`
+  deliberately not rotated: out-of-band issue work.
+
 - 2026-08-09 Signal-time terminal restore (`sig-leave`, closes #12): the port
   handled only SIGHUP and SIGTERM, so SIGINT and SIGQUIT killed the process with
   tcell still holding the tty, leaving the user at a shell with no echo. All