test: plan layer safety e2e tests (68 tests)

Comprehensive user-perspective tests for the plan refinement changes:

## Phrase Classification (42 tests)
- 14 approval cases ("looks good", "proceed", "lgtm", etc.)
- 12 rejection cases ("no", "stop", "reject", etc.)
- 8 pure refinement cases (feedback without approval/rejection signals)
- 8 tricky edge cases:
  - "yes, but change X" → refine (qualifier overrides approval)
  - "no, I mean yes" → reject (rejection takes priority)
  - "I know this looks good" → approve ("know" ≠ "no" via word boundary)
  - Unicode input, multiline, special characters, empty string

## Non-Plan Agent Safety (3 tests)
- Plan variables initialized before loop
- Refinement block guarded by `agent.name === "plan"`
- Plan file detection guarded by agent check

## Session Agent Name Fix (2 tests)
- `sessionAgentName` set before early break conditions
- `agent_outcome` telemetry uses `sessionAgentName`

## Revision Cap (4 tests)
- Cap enforced at >= 5
- Synthetic message communicates limit to LLM
- Telemetry emits "cap_reached"
- Synthetic message doesn't persist to DB

## Adversarial (4 tests)
- 70k char input doesn't crash
- Unicode, special chars, multiline handled

## Import Safety (2 tests)
- post-connect-suggestions is self-contained (no heavy imports)
- Progressive suggestion dedup works correctly

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: anandgupta42