fix(agy): report the failure reason instead of echoing the answer

[junmo-kim]

Problem

On an agy session, a turn that ends with a non-SUCCESS result envelope
delivers its answer twice: once as the streamed agy_message, then again as an
error event right below it — the whole answer rendered as plain text with a ⚠
prefix.

The driver takes the failure text from response:

if (event.status !== 'SUCCESS') {
    resultFailure = event.response?.trim()
        || `agy turn failed: ${event.status}`;
}

But response is the answer field. agy fills it from the completed buffer
regardless of status, so on a failed turn it holds the same answer that already
went out through the delta stream. The reason for the failure is reported
separately, in error, which the parser did not read at all.

Measured on agy 1.1.13 (--print-timeout 15s against a long prompt):

{"event":"result","result":{"conversation_id":"…","status":"ERROR","response":"","error":"timeout waiting for response","duration_seconds":2.4,…}}

response is empty and the reason lives in error. On a turn that produced
text before failing, response carries that text instead — which is what ends
up duplicated in the chat.

Change

• agyNdjsonParser.ts: carry error through the result envelope (added to
  ResultEnvelope and to the result stream event; null when absent).
• agyHeadlessDriver.ts: read the failure text from error rather than
  response. The existing agy turn failed: <status> fallback still applies
  when the envelope carries no reason, so a failed turn stays visible.

The SUCCESS branch is untouched.

Tests

• parser: an envelope with error surfaces it; the existing SUCCESS/FAILURE
  cases now assert error: null.
• driver: a failed envelope whose response holds the answer reports the
  error text and never emits the answer as an error; a failed envelope with no
  error falls back to the status; a timeout envelope surfaces
  timeout waiting for response.

bun run test:cli and bun run typecheck:cli pass.

Note

#1629 fixes the same duplication on the SUCCESS path (where the guard comparing
the streamed text with the envelope misfires on mangled deltas). This one is the
failure path, which that guard never reaches. The two touch adjacent hunks of the
same file but do not depend on each other in either order.

AI disclosure per CONTRIBUTING: written with Claude Code (Opus 5).

[github-actions]

Findings

• No reportable issues found.

Summary

• Review mode: initial. Reviewed the full four-file diff; no correctness, security, regression, data-loss, performance, or maintainability issue met the 80% confidence threshold. A formal agy result-envelope schema was not found in repo/docs, so compatibility remains dependent on the observed CLI envelope shape.

Testing

• Not run (automation; PR code was not executed). Static inspection only; GitHub test and integration checks were still in progress at review time.

HAPI Bot