DIVE-3786: stall alert stops naming an untested mechanism; re-arm kick verifies its own submit

[5dive-bot]

DIVE-3786 — stop reporting an untested mechanism, and make the re-arm kick observe its own submit

A customer reported seeing this constantly, from a codex box:

can't respond right now — listen loop wedged / agent is idle outside wait_for_message and won't re-arm

Defect 1 — that string was the else-branch

detectStallCause() captures the agent's tmux pane, takes the last 4000 chars, and tests exactly two
regexes (quota exhausted, auth expired). Everything else returned listen loop wedged / agent is idle outside wait_for_message and won't re-arm — a named mechanism, in the customer's chat, from a
function that reads neither the listen loop, nor wait_for_message state, nor re-arm state. Its
frequency was evidence about the classifier's coverage, not about the listen loop.

Fix: the unmatched branch now returns not responding — cause unknown plus the last 3 non-empty
lines of the agent's screen (paneTailSummary), so the owner sees what the agent is actually sitting
on. listen loop wedged no longer exists in the repo.

Defect 2 — the recovery path broke itself, which is why it fired constantly

Root cause found on the customer box: the agent had finished its turn cleanly and a line was sitting
unsubmitted in the codex composer. kickListenLoop() was:

send-keys -l REARM_KICK_TEXT   // -l APPENDS to whatever is already in the composer
setTimeout(400) -> send-keys Enter, callback () => {}   // result discarded

send-keys -l appends, so each kick concatenated onto the stranded line instead of replacing it; and
the Enter was fire-and-forget, directly under a comment noting the TUI "occasionally drops an Enter".
Three unnoticed failures (STALL_ESCALATE_AFTER) then fired the owner alert. Self-perpetuating by
construction: the recovery mechanism used the delivery path that was broken.

Fix, per attempt: C-u (clear composer) → send-keys -l <text> → 400 ms → capture pane →
send-keys Enter (own call) → 700 ms → capture pane → assert the pane changed. A landed Enter
clears the composer and starts a turn, so byte-identical panes mean the keystroke went nowhere. One
retry, then rearmSubmitFailed = true. A pane we cannot read counts as unconfirmed, not as success.

rearmSubmitFailed is the only wedged state we actually measure, so it is the only one the alert
names: "stuck at the input prompt". clearStallAlert() resets it on recovery.

Also

sendStallAlert() writes <STATE_DIR>/last-stall-pane.txt (timestamp, agent, cause, full pane) next
to the existing stamp, so the next occurrence is diagnosable on the box rather than from a forwarded
screenshot.

Applied to all three forks — telegram-codex, telegram-grok, telegram-agy — the code was
identical in each, so a codex-only fix would have left the same string shipping on grok and
antigravity boxes.

Evidence

• C-u clears a live codex composer — probed against agent-andy (codex TUI, 2026-08-28): typed
  a probe string, composer showed it; C-u returned the composer to its placeholder hint. No Enter
  was sent, nothing was submitted.
• The pane-change check fires in both directions — real tmux session running cat: text + Enter →
  pane changed (green when a submit lands); the same pane sampled twice with no input → byte-identical
  (so the check can genuinely FIRE, not merely fail to false-alarm). C-u there took a stranded line
  from 1 occurrence to 0.
• Full suite: 1060 pass / 0 fail across 37 files on the patched tree.
• New guard test/dive3786-honest-stall-cause.test.ts — 27 assertions across the 3 forks.
  Negative control: all 27 FAIL against origin/main's server.ts; all 27 pass on this branch.
• bun build --external '*' parses all three patched servers. No repo-root tsconfig.json, so there
  was no typecheck arm to run — not claimed.

Deliberately out of scope, follow-up owed

5dive-agent-start (repo 5dive-cli, grok :1708 / antigravity :1733 cold-start kicks) has the same
unverified -l → sleep 1 → Enter two-step. Left out: different repo, different ship path
(deliberate release-cut, and the boot path owes a smoke this host cannot run), it is the boot path
for every agent, and defect 1 cannot bite there — the cold-start composer is empty by construction,
single-fire, so only the unverified-Enter half applies.

Separate finding while reading it: both CLI _kick strings contain "keep looping" and "call it
again immediately" — verbatim the model-facing phrases test/rearm-loop-regression.test.ts bans as
the DIVE-1180 idle-token-burn cause. That guard scans only this repo, so the CLI copy is unguarded.
Worth its own row.