spritzer 0.4.1 -> 0.5.0, and the turn completes every time

[jhgaylor]

spritzer:0.4.1 was released 2026-07-14. The two fixes this repo needed merged
on 08-03 and 08-04. latest and 0.4.1 were byte-identical manifests, and
the release workflow only fires on a v* tag — so nothing published carried
either fix, and every deployment kept reproducing a bug that was already fixed
on main. spritzer 0.5.0 is now
cut; this moves to it.

What closes

With fountain#603 already in the v0.6.1 pin, three fixes across two repos:

fountain#603	a lost stdin write crashed the ConversationServer; the restart found the sandbox ready, reattached, and orphaned the turn
spritzer#19	an unupgraded GET on the exec path answered 426 instead of the session list — the call fountain makes before deciding to reattach
spritzer#20	an unrecognised command now holds its exec session open until stdin EOF, so the prompt reaches a process that is still there

Measured

34 of 34 conversations completed at fountain v0.6.1 + spritzer 0.5.0,
including batches fired back to back — the pacing that used to fail most. Zero
crashes, zero reattaches, zero 426s, zero :command_exited.

	0.4.1	0.5.0
completed	13 of 30	34 of 34
failed :command_exited	17	0

And the prompt now makes the round trip, which is more than an exit code:

event: output  claude --dangerously-skip-permissions --print --verbose …
event: output  Reply with the single word: fountain
event: stage   turn  done  {"exit_code":0}

The gate stops hedging

It asserted a result, then a pairing, then only observed which ending it saw.
Each retreat was right at the time — the ending was a race. It is not one now,
so both former outcomes become failures, and each names the pin that must have
moved:

✗ the turn was orphaned behind a reattach — fountain#603 or spritzer#19
  regressed, or the pin rolled back
✗ the runtime exited before the prompt was written — spritzer#20 regressed,
  or spritzerImage rolled back below 0.5.0

verify-conversation goes back to one assertion for both planes: exit 0 and
output streamed.

Docs

Completing a turn moves to Verified, and Does not work is now empty —
with a line saying why, so an empty section does not read as an oversight.

That the local turn is the echo rather than a model stays as a qualifier on the
Verified row and in data-plane.md's what it will never prove, where it
belongs. It is not a defect of this deployment: whether a model replies depends
on a real data plane and a tenant's own inference credentials, which is outside
what this repo deploys.

data-plane.md keeps the history in a collapsed note rather than deleting it.
This was described wrongly three times — as an arm64/amd64 split, then as a
dispatch race "decided by machine speed" from 5-of-5 and 2-of-2 samples, then
as "no turn finishes, by upstream design" while v0.6.0 was completing 13 of 30.
Every one was a small sample of a coin flip read as a property, and that is
worth leaving on the page.

Verification

just e2e green from nothing at these pins, with the strict gate:

✓ registered and verified e2e-1786080723@example.com
✓ e2e-1786080723@example.com is admin, audit-recorded
── the conversation gate ──
✓ the turn completed, prompt echoed back — the echo, not a model
✓ every seam validated
✓ e2e: every documented claim held, from nothing, and the cluster is gone.

just check and just site clean.

Supersedes #105, which described a race this makes historical — I will close it.

🤖 Generated with Claude Code