fix(backend): teach LlmModel about OpenRouter Anthropic aliases + dry-run sim default Haiku OR-slug (#13177)

### Why / What / How

**Why:** PR #13171 (SECRT-2368) flipped the dry-run simulator default to
`LlmModel.CLAUDE_4_5_HAIKU.value` = `"claude-haiku-4-5-20251001"`. That
broke every dry-run on dev because the LLM-simulation path
(`_call_llm_for_simulation`) hits OpenRouter's OpenAI-compat endpoint
(`/api/v1/chat/completions`), which only accepts canonical
`<vendor>/<model>` slugs and returns HTTP 400 `"<id> is not a valid
model ID"` for direct-Anthropic snapshot IDs. User-visible: `[SIMULATOR
ERROR — NOT A BLOCK FAILURE] Failed after 5 attempts: Error code: 400 -
'claude-haiku-4-5-20251001 is not a valid model ID'`.

**Verified empirically against OpenRouter:**

| Slug | `/v1/chat/completions` (OpenAI-compat) | `/v1/messages`
(Anthropic-compat) |
|---|---|---|
| `claude-haiku-4-5-20251001` (snapshot) | ❌ 400 not valid | ✅ accepted
|
| `anthropic/claude-haiku-4-5` (OR slug) | ✅ accepted | ✅ accepted |
| `anthropic/claude-haiku-4-5-20251001` (prefix + snapshot) | ❌ 400 not
valid | — |

So the right default for the simulator is the **OpenRouter slug**
(`anthropic/claude-haiku-4-5`), and we need `LlmModel` to resolve that
string back to the existing `CLAUDE_4_5_HAIKU` enum member so
`OrchestratorBlock.Input.model` validation still passes.

**What:**
- Add `_OPENROUTER_ALIASES: dict[str, str]` in `backend/blocks/llm.py`
mapping the three Claude 4.5 OpenRouter slugs
(`anthropic/claude-haiku-4-5`, `anthropic/claude-opus-4-5`,
`anthropic/claude-sonnet-4-5`) to their direct-Anthropic snapshot enum
values. The newer 4.6/4.7 entries don't carry the snapshot suffix, so
the existing prefix-strip path already covers them.
- Extend `LlmModel._missing_` to consult `_OPENROUTER_ALIASES` first,
then fall back to the existing provider-prefix strip.
- Change `_DEFAULT_SIMULATOR_MODEL` (in `backend/executor/simulator.py`)
and `ChatConfig.simulation_model` default (in
`backend/copilot/config.py`) to the bare string
`"anthropic/claude-haiku-4-5"`. Both files drop their now-unused
`LlmModel` import.
- Update `TestDefaultSimulatorModel` to pin the new default + assert it
resolves to `LlmModel.CLAUDE_4_5_HAIKU` via the alias map, plus a `"/"
in default` guard so a future revert to a snapshot ID trips at unit-test
time.
- New `TestLlmModelMissingHandler` covering: alias-map resolution,
prefix-strip fallback, and unknown-slug raising.

**How:**
- **LLM-simulation path** (`_call_llm_for_simulation`): sends
`"anthropic/claude-haiku-4-5"` straight to OpenRouter's OpenAI-compat
endpoint via `get_openai_client(prefer_openrouter=True)` — accepted.
- **Orchestrator dry-run path** (`prepare_dry_run` injects
`"anthropic/claude-haiku-4-5"` as `model` on `OrchestratorBlock.Input`):
Pydantic calls `LlmModel("anthropic/claude-haiku-4-5")`, hits
`_missing_`, alias map returns `CLAUDE_4_5_HAIKU`. The orchestrator's
Anthropic SDK then sends the enum's `.value`
(`"claude-haiku-4-5-20251001"`) to OpenRouter's Anthropic-compat
endpoint, which accepts both formats — also fine.

**Why not duplicate enum entries?** Per reviewer feedback ("this is
weird having multiple models"): adding `CLAUDE_4_5_HAIKU_OPENROUTER`
alongside the existing direct-Anthropic entry would duplicate metadata +
cost lookups for the same model and pull the OpenRouter/direct routing
distinction into the enum itself. The alias map keeps `LlmModel` as one
entry per model while still letting both identifiers resolve through the
validator.

### Changes 🏗️

- `_OPENROUTER_ALIASES` map + extended `_missing_` in `LlmModel`.
- Dry-run simulator default flipped from Haiku snapshot ID to Haiku
OpenRouter slug.
- New unit tests for the alias-resolution path; updated
`TestDefaultSimulatorModel` pin.
- No `LlmModel` enum members added or removed → no `llm.md` regen
needed.

### Checklist 📋

#### For code changes:
- [x] I have clearly listed my changes in the PR description
- [x] I have made a test plan
- [x] I have tested my changes according to the test plan:
- [x] `poetry run pytest
backend/executor/simulator_test.py::TestDefaultSimulatorModel
backend/executor/simulator_test.py::TestPrepareDryRun
backend/copilot/tools/test_dry_run.py::test_prepare_dry_run_orchestrator_block`
— all green
- [x] `poetry run pytest
backend/blocks/test/test_llm.py::TestLlmModelMissingHandler` — 3 new
tests, all green
  - [x] Empirical OpenRouter probes recorded above
- [x] Direct e2e check: `LlmModel("anthropic/claude-haiku-4-5") is
LlmModel.CLAUDE_4_5_HAIKU` ✓; `LlmModel("anthropic/claude-opus-4-7") is
LlmModel.CLAUDE_4_7_OPUS` ✓
- [x] `poetry run ruff format` + `poetry run ruff check` clean on all
changed files

---------

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