…ty flip + anthropic runtime dep

Closes Phase 2 Feature 5's biggest user-facing milestone. Joshua
DMs Kora → Kora reasons (real LLM call, system-prompt-driven,
context-aware, cost-ladder-aware) → Kora replies with the
reasoning output. Echo path is gone.

## Three PM rulings baked in

**Ruling 1 — Credential cascade FLIPPED to OAuth-first.**

ST1 shipped `KORA_ANTHROPIC_API_KEY → CLAUDE_CODE_OAUTH_TOKEN`
order. ST2 flips to `CLAUDE_CODE_OAUTH_TOKEN → KORA_ANTHROPIC_API_KEY
→ fail-CLOSED`. Reasoning: Joshua's Max 20x + post-May-15 SDK
billing split route the $200/mo Agent SDK pool via the OAuth path.
OAuth = production; API key = test/dev escape hatch. Engine's
`_auth_mode` test updated to assert OAuth wins when both are set.

**Ruling 2 — `anthropic` to runtime deps.**

Same lesson as KR-SLOWAPI-DEP-FIX + aiosmtplib promotion: the
reasoning_engine listener imports anthropic unconditionally at
boot, so it belongs in the core dependencies block, not under
`[anthropic]` or `[web]` extras. Both prior locations removed;
single declaration in core deps.

**Ruling 3 — §4 PM-opens locked from ST1.**

Q1 system prompt first-draft (composed in ST1); Q2 10-turn context;
Q3 NO retry on 5xx; Q4 PAUSED state gating. All carry through ST2.

## New module

**`kora_cli/listeners/reasoning_engine_listener.py`** (~135 lines):

- `ReasoningEngineListener` wrapping the engine in the
  DaemonCoordinator lifecycle. Startup constructs
  `AnthropicReasoningEngine`; **construction failure RE-RAISES**
  so the coordinator aborts boot per the spec's "engine startup
  failure → daemon fails-CLOSED" requirement (a daemon that
  can't reason can't fulfill its primary purpose; better to abort
  loudly than ship a fallback-only daemon).
- Module-level `current_reasoning_engine()` accessor mirroring
  `current_pool()` from KR-MCP-CONSUMPTION ST1 — cross-cutting
  read from `SlackDMHandler` without import coupling.
- `_set_singleton` / `_clear_singleton` private helpers (same
  shape as the MCP pool listener).
- Shutdown closes the engine's HTTP client best-effort; per-listener
  timeout (10s) caps the wait.
- Registered via `register_daemon_listener("reasoning_engine", _factory)`
  at import time; wired into `kora_cli/listeners/__init__.py`.

## Handler swap

**`kora_cli/handlers/slack_dm_handler.py`** (+350 lines):

- New optional `reasoning_engine` constructor arg (test injection).
  Production resolves the daemon singleton via
  `current_reasoning_engine()` at reply time.
- Method `_send_echo_reply` (name kept for diff hygiene) body
  swapped: builds `ConversationContext` via
  `load_slack_dm_context()`, constructs `IncomingMessage`, calls
  `engine.respond(...)`, projects `ResponseResult` into reply
  text + JSONL meta.
- **Canned fallback** `"Kora is currently unable to respond;
  operator notified."` sent when:
  - No engine resolvable (test path / partial daemon boot)
  - `ResponseResult.error` set (cost_ladder_halted /
    operational_state_paused / sdk_5xx / ...)
  - Engine itself raises (defensive — recorded as
    `engine_exception:<class>`)
  - Engine returns empty text on success (defensive —
    `empty_response_text`)
- Reply failure does NOT crash the inbound handler — Slack
  always gets 200 OK (preserves the KR-FEAT-SLACK-DM ST2
  belt-and-suspenders guarantee).

## JSONL outbound schema extended

5 new optional fields on outbound entries (backwards-compatible —
pre-ST2 entries simply omit them):

- `model_used`: e.g. `"claude-opus-4-7"`
- `input_tokens` / `output_tokens`: from SDK usage
- `reasoning_duration_ms`: engine-side wall-clock
- `reasoning_error`: stable error code (None on success)

Each field is written only when non-None, so canned-fallback
entries that lack reasoning data stay lean.

## Cost-ladder write integration

After a successful reasoning call (NOT canned fallback), the
handler builds `CanonicalUsage(input_tokens, output_tokens)`
and calls `holder.record_inference(canonical_usage, model_name,
provider="anthropic")`. Fail-soft layers:

- No cost holder initialized → silently skip (test paths)
- Zero tokens → skip (no real inference happened)
- `record_inference` itself is fail-soft per its docstring
  (pricing-lookup miss accumulates 0)

K-DG verified literal API: `record_inference(canonical_usage, *,
model_name, provider, base_url)` per
`agent/cost_state_holder.py:272`; `CanonicalUsage` shape per
`agent/usage_pricing.py:30` (input_tokens / output_tokens /
cache_read_tokens / cache_write_tokens / reasoning_tokens /
request_count / raw_usage).

## Tests (57 new, 332 total all passing)

**`test_reasoning_engine_listener.py`** (8 tests):
- current_reasoning_engine() None pre-startup
- Startup with injected engine sets singleton
- Shutdown clears singleton + closes engine
- Shutdown safe when engine.close() raises
- Shutdown safe without prior startup (no-op)
- **Production startup re-raises on misconfig** (fail-CLOSED
  asserted — both credential envs unset → coordinator boot abort)
- Registered in LISTENER_REGISTRY as "reasoning_engine"
- _set_singleton / _clear_singleton helpers tested

**`test_slack_dm_reply.py`** updated/extended (15 ST2 tests):
- ST1 echo-format tests rewritten to assert engine's response
  text (not echo construction)
- `test_oauth_wins_over_api_key_when_both_set` — Ruling 1 verified
- Engine unavailable → canned fallback + reasoning_error="engine_unavailable"
- Engine returns error (cost_ladder_halted / operational_state_paused)
  → canned fallback + recorded error code
- Engine raises exception → canned fallback +
  reasoning_error="engine_exception:<class>"
- Empty engine text on success → canned fallback + "empty_response_text"
- Engine receives correct IncomingMessage metadata (source /
  channel_id / thread_ts / user_id / event_ts)
- **Cost-ladder write called on success** with correct
  CanonicalUsage + model_name + provider="anthropic"
- Cost-ladder write SKIPPED on canned fallback (no real inference)
- Cost-ladder gracefully skips when holder uninitialized

## §5 ship checklist

- [x] Base `feature/phase2-upgrades`
- [x] Title `feat(kora): KR-FEAT-AI-RESPONSE-LOOP ST2 — handler swap + cred priority flip + anthropic runtime dep`
- [x] §4 PM-opens carried from ST1 (all defaults locked)
- [x] Credential cascade flipped OAuth-first per Ruling 1
- [x] anthropic moved to runtime deps per Ruling 2
- [x] API key NEVER logged (existing ST1 diverse-failure test
      still holds — credential value never in any error / log path)
- [x] Cost-ladder write integration verified (new test)
- [x] PAUSED state respected (engine refuses; handler sends canned)
- [x] Engine startup failure → daemon fails-CLOSED (re-raise asserted)
- [x] Engine failure paths → canned fallback (5 distinct paths tested)
- [x] K-DG: literal field names pasted from grep:
      `record_inference(canonical_usage, *, model_name, provider, base_url)`,
      `CanonicalUsage` shape, `get_cost_holder()` accessor,
      `_HOLDER` reset helper
- [x] Tests pass locally (**332/332** across full suite)

## What's next

After ST2 merges, **Phase 2 Feature 5 closes**. Joshua DMs Kora,
Kora reasons (real Anthropic API call against the Max plan pool,
cost-ladder-aware model selection, operational-state-respecting,
10-turn-context-aware), Kora replies intelligently.

PM picks next bucket — candidates:
- **KR-FEAT-AGENTIC-REASONING**: Kora can call kora__* tools inside
  her reasoning (look up ledger, check sea_tickets) for richer
  responses
- **KR-MCP-SEND-TOOLS**: expose kora__send_slack_dm + kora__send_email
  via /mcp for agent-driven sends
- **Substrate-backed conversation memory**: persist context to
  IsoKron substrate, not just JSONL

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>