fix(coding-agent): resume a provider-error-blocked goal on the next user message

[code-yeongyu]

Problem

A goal blocked by a terminal provider error stayed blocked forever. agent_end writes
blocked with reason provider error ended the turn (retries exhausted) when
AgentEndEvent.willRetry is false, but MECHANICAL_CONTINUATION_BLOCKS only listed the
cap, repetition, and length guards. GoalDirectInputLifecycle.onDisposition reactivates a
blocked goal on accepted direct input only for mechanical blocks, so the user's next
message — the natural retry signal — did nothing, and the blocked notice never said how to
recover.

Change

• continuation-recovery.ts exports PROVIDER_ERROR_BLOCKED_REASON and adds it to
  MECHANICAL_CONTINUATION_BLOCKS.
• index.ts writes that shared constant instead of repeating the literal reason, and
  appends continuationCapRecoveryHint(...) to the blocked notice so the warning ends with
  Send any message to resume.
• direct-input-lifecycle.ts needed no change: the reactivation path already existed; the
  provider-error reason now flows through it.

Intentional blocks stay non-recoverable — user interrupted the turn and model-declared
update_goal blocks encode a decision to stop, and a new regression test pins that.

Prior art check: codex ext/goal (runtime.rs stop_active_goal_for_turn) also marks a
turn error Blocked, but its resume is user/system controlled with no user-message
auto-resume — so this is a deliberate senpi-side policy improvement, not a parity port.

Evidence

RED (before the production change), both criteria failing for the right reason:

FAIL goal-extension.test.ts > resumes a provider-error-blocked goal when the user sends a new message
  expected { status: 'blocked' } to match object { status: 'active' }
FAIL goal-cap-recovery-guidance.test.ts > classifies a terminal provider error as prompt-recoverable
  expected undefined to be 'provider error ended the turn (retries exhausted)'

GREEN:

Test Files  5 passed (5)
     Tests  118 passed (118)

(goal-cap-recovery-guidance, goal-extension, issue-447-goal-continuation, goal-modules, goal-store)

Real-surface run against the actual store + lifecycle modules in a temp goal dir:

1. after terminal provider error -> blocked / provider error ended the turn (retries exhausted)
2. after a new user message -> active
RESULT: PASS
cleanup: removed /var/folders/.../senpi-goal-qa-bwf95O

Root npm run check: clean.

Docs

goal/changes.md gets a dated fork entry with merge-conflict zones; goal/AGENTS.md
CONTINUATION POLICY now states the mechanical/intentional split.

---

Summary by cubic

Treats terminal provider errors as mechanical blocks so a new user message resumes the goal. The warning now includes “Send any message to resume.”

• Bug Fixes
  • Added PROVIDER_ERROR_BLOCKED_REASON to MECHANICAL_CONTINUATION_BLOCKS, enabling automatic reactivation via existing direct-input handling.
  • Updated index.ts to use the shared reason and append continuationCapRecoveryHint(...) to the warning.
  • Intentional blocks (user interrupt, model update_goal) remain non-recoverable; tests and docs updated.

^{Written for commit c3f2997. Summary will update on new commits.}

[code-yeongyu]

Rebased onto main after #687 merged.

#687 and this PR are complementary halves of the same recovery story and both landed the same day:

• fix(goal): prompt to resume blocked goals on restart #687 — restart resume prompt: a new session loading a stopped (paused | blocked) goal now prompts Resume blocked goal?. Requires a restart and a user click.
• this PR — in-session resume: a goal blocked mid-run by a terminal provider error resumes on the next user message, with no restart and no prompt, because the reason is now classified mechanical.

Without this PR, a provider outage mid-session still stranded the goal until the user quit and relaunched senpi.

Rebase touched goal/changes.md only (both PRs prepended a dated entry); both entries are preserved and cross-referenced. goal/AGENTS.md merged cleanly — #687's RESTART RESUME PROMPT section and this PR's CONTINUATION POLICY sentence coexist.

Re-verified on the new base:

full goal surface (7 files)   142 passed (142)
npm run check                 exit 0

Note on the earlier red run: Test (coding-agent 2/3) failed on test/mcp/catalog-cache.test.ts ("refreshes eager servers from the live catalog when a valid warm cache differs"). That file is untouched by this diff — which is confined to goal/ — and passes locally (6/6). Flaky, unrelated to this change.