fix(backend): unblock dry-run Orchestrator + Haiku 4.5 sim default (#13171)

### Why / What / How

**Why:** The Agent Simulator (dry-run) routes the Orchestrator block to
*real* execution with a platform-supplied "cheap" model. Two problems
with the pre-PR state:

1. The platform default was `"google/gemini-2.5-flash-lite"` — a string
literal that was **not a member of the `LlmModel` enum**. Since
`OrchestratorBlock.Input.model` is typed as `LlmModel`, Pydantic input
validation rejected every dry-run that hit the Orchestrator path.
2. The "cheap" model choice (Flash-Lite) was justified as
"shape-matching only", but Orchestrator dry-runs aren't shape-matching —
they pick tools and emit a finish message. A substitute that's too weak
gives an unfaithful preview of what the real run would do.

**What:**
- Register the missing `google/gemini-2.5-flash-lite` slug as a
first-class `LlmModel` member (so any deployment that prefers Flash-Lite
via `CHAT_SIMULATION_MODEL` keeps working — fixes the validation crash).
- Switch the platform default from Flash-Lite to **Claude Haiku 4.5** so
the dry-run preview better matches the real run's tool-use decisions.
- Tie both default constants (`simulator.py:_DEFAULT_SIMULATOR_MODEL`
and `copilot/config.py:simulation_model`) to
`LlmModel.CLAUDE_4_5_HAIKU.value` at import time, so any future rename
or drift breaks at module load instead of silently failing input
validation.
- Update the `TestDefaultSimulatorModel` pin to match the new default,
and add an `LlmModel(_DEFAULT_SIMULATOR_MODEL) is
LlmModel.CLAUDE_4_5_HAIKU` assertion so the bug class behind SECRT-2368
(a default that isn't an enum member) trips at unit-test time.
- Tighten the two `prepare_dry_run(OrchestratorBlock, …)` tests with the
same `LlmModel`-parseability predicate `OrchestratorBlock.Input`
enforces at runtime.

**How:**
- `backend/blocks/llm.py`: add `GEMINI_2_5_FLASH_LITE =
"google/gemini-2.5-flash-lite"` enum entry + matching `ModelMetadata`
(1M context, 64K max output, tier 1, OpenRouter).
- `backend/data/block_cost_config.py`: matching entries in `MODEL_COST`
(tier 1) and `TOKEN_COST` (`input=15, output=60` — OpenRouter's
$0.10/$0.40 × 1.5× margin × 100 credits/$, matches the existing
`GEMINI_2_0_FLASH_LITE` formula).
- `backend/executor/simulator.py`: import `LlmModel` at top, set
`_DEFAULT_SIMULATOR_MODEL = LlmModel.CLAUDE_4_5_HAIKU.value`, update the
comment to reflect the decision-quality rationale.
- `backend/copilot/config.py`: import `LlmModel` at top, set
`simulation_model` default to `LlmModel.CLAUDE_4_5_HAIKU.value`, update
the description.
- Tests: move `from backend.blocks.llm import LlmModel` to top-level per
`backend/AGENTS.md`; update the `TestDefaultSimulatorModel` pin + add an
`LlmModel`-parseability check on the default; add a parseability
assertion on the `prepare_dry_run` orchestrator tests.
- `docs/integrations/block-integrations/llm.md`: regenerated via `poetry
run python scripts/generate_block_docs.py` so `check-docs-sync` passes
with the new enum member.

**Routing note (why Haiku-via-OpenRouter is safe):** `prepare_dry_run`
injects an OpenRouter API key as the credential, and `OrchestratorBlock`
dispatches on the **credential**'s provider (`orchestrator.py:1647`),
not the model's metadata provider. So Haiku 4.5 routes through
OpenRouter's Anthropic-compat endpoint regardless of
`LlmModel.CLAUDE_4_5_HAIKU.metadata.provider == "anthropic"`. No change
to provider routing.

### Changes 🏗️

- New `LlmModel.GEMINI_2_5_FLASH_LITE` enum entry + metadata + cost
lookups.
- Dry-run simulator default flipped from Flash-Lite to Claude Haiku 4.5.
- Both default constants tied to the `LlmModel` enum at import time.
- `TestDefaultSimulatorModel` pin updated; `LlmModel`-parseability
assertion added on both the pin and the two `prepare_dry_run`
orchestrator tests.
- Regenerated `llm.md` for the new enum member.

### 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/copilot/tools/test_dry_run.py::test_prepare_dry_run_orchestrator_block`
passes
- [x] `poetry run pytest
backend/executor/simulator_test.py::TestPrepareDryRun` passes (5 tests)
- [x] `poetry run pytest
backend/executor/simulator_test.py::TestDefaultSimulatorModel` passes (2
tests — value pin + `LlmModel` parseability)
- [x] Direct end-to-end check: `prepare_dry_run(OrchestratorBlock(),
{…})` injects `"claude-haiku-4-5-20251001"` and
`OrchestratorBlock.Input(...)` validates to `LlmModel.CLAUDE_4_5_HAIKU`
without error
- [x] Switching `CHAT_SIMULATION_MODEL=google/gemini-2.5-flash-lite`
still works (the enum entry was added for exactly this configurable
case)
- [x] `poetry run ruff format` + `poetry run ruff check` clean on all
changed files