feat(backend/copilot): switch main client between OpenRouter and Anthropic-direct (#13034)

## Why

Today copilot routes every LLM call through OpenRouter. We want the SDK
+ baseline paths to talk directly to Anthropic (api.anthropic.com) so we
avoid the OR proxy hop for models we run on Anthropic anyway, while
keeping non-Anthropic helpers (title generation, future builder helpers)
on OR — they need providers Anthropic can't serve.

The plumbing for the SDK side mostly existed (the SDK env builder
already had a direct-Anthropic mode), but the baseline path and the
title client both passed through the same OpenRouter-pinned
`AsyncOpenAI` singleton, so flipping `CHAT_USE_OPENROUTER=false` would
silently break aux calls.

## What

- **Default flipped** — `CHAT_USE_OPENROUTER` now defaults to `false`
(direct-Anthropic main path). OpenRouter is opt-in via env var.
- **Config split** — added `aux_api_key` / `aux_base_url` /
`direct_anthropic_api_key` fields to `ChatConfig`, plus
`main_client_credentials` / `aux_client_credentials` /
`aux_uses_openrouter` / `aux_provider_label` / `openrouter_active` /
`effective_transport` properties so call sites read intent instead of
re-deriving URL checks. URL host comparison uses `_host_matches()`
(urlparse-based, fixes a CodeQL incomplete-substring sanitization).
- **Two client functions** — `_get_main_client()` (transport-aware:
OpenRouter or `https://api.anthropic.com/v1/`) and `_get_aux_client()`
(defaults to OR; falls back to main creds for single-key deployments).
`_get_openai_client` kept as a back-compat alias.
- **Boot validators** — `_validate_sdk_model_vendor_compatibility`
(non-Anthropic slug + direct mode → boot fail) and
`_validate_aux_client_for_direct_main` (catches `aux_base_url` without
key, subscription-mode + no aux key + no direct key, non-Anthropic title
model paired with Anthropic-pointed aux client).
- **Title generation** uses the aux client; provider label resolves to
`open_router` / `anthropic` / `openai` via `aux_provider_label`. Title
model is normalized for the Anthropic OpenAI-compat endpoint (strip
`anthropic/` prefix and dots) when aux lands on Anthropic.
- **Baseline path** normalizes the model slug for the active transport
(`model_normalize.py`), drops OR-only `usage.include`, `reasoning`, and
`stream_options` (Anthropic compat 400s on these) when not on OR, swaps
to native `thinking` extra_body in direct mode, and clamps
`budget_tokens` + `max_tokens` against the model's published max output
(Anthropic compat requires `max_tokens > budget_tokens`).
- **Anthropic rate card** — vendored `anthropic_rates.json`
(LiteLLM-derived, refreshed weekly via cron).
`compute_anthropic_cost_usd()` computes USD when OR's `usage.cost`
extension is absent (direct mode), with TTL-aware cache-write
multipliers (1.25× input for 5m, 2.0× for 1h) and a 0.1× cache-read
multiplier. Cache token buckets subtracted from `prompt_tokens` to keep
the three buckets disjoint. Falls back to opus-4-1 rates for unknown
Anthropic slugs (over-bill rather than mis-bill). Drops `litellm`
dependency entirely (snyk fix — repeated auth-bypass / password-hash
CVEs).
- **Cost recovery safety net** — when the upstream stream finishes
without a `usage` chunk in direct mode (truncated SSE / mid-stream
disconnect), recover cost from the rate card with tiktoken-estimated
tokens so the rate-limit charge can't be bypassed. OR mode skips
recovery (OR markup ≠ raw Anthropic pricing).
- **Shared normalizer** — `_normalize_model_name` extracted into
`copilot/model_normalize.py` so SDK and baseline share one rule. SDK
keeps the symbol as a thin wrapper passing its own `config` reference,
so existing monkeypatch fixtures keep driving the decision.
- **Shared helper** — `_extract_cache_creation_tokens` lives in
`token_tracking.py` and is consumed by both `service.py` and
`baseline/service.py`.

## How

Switching is a deploy-wide env flip:

```bash
# Direct Anthropic (new default)
CHAT_USE_OPENROUTER=false
ANTHROPIC_API_KEY=sk-ant-...
# (optional — defaults to OPEN_ROUTER_API_KEY for aux client)
CHAT_AUX_API_KEY=sk-or-v1-...

# OpenRouter (opt-in)
CHAT_USE_OPENROUTER=true
```

No LaunchDarkly flag — the existing `CHAT_USE_OPENROUTER` env var is
enough for a deployment-wide cutover. Per-tenant rollout can layer LD on
top later if needed.

USD cost tracking is preserved: SDK direct mode uses the CLI's
`total_cost_usd` (rate-card accurate for Anthropic models); baseline
direct mode uses the rate-card lookup, recorded with
`provider="anthropic"`. Aux calls keep using OR's `usage.cost` extension
when on OR; when aux falls back to Anthropic, cost is computed via the
rate card at the title-cost recorder.

### Test plan

- [x] `pytest
backend/copilot/{config,anthropic_rate_card,model_normalize,service_unit,baseline/service_unit}_test.py`
+ `sdk/{service,service_helpers,openrouter_cost,env}_test.py` — all
green locally
- [x] New tests cover stream_options gating, direct-mode provider label,
cost recovery on missing usage chunk
- [x] Smoke verified `ChatConfig.main_client_credentials` returns
`(or-key, or-url)` in OR mode and `(anthropic-key,
https://api.anthropic.com/v1/)` in direct mode
- [x] /pr-test E2E on native stack — both transports, both modes,
paywall resumption, title generation, thinking-budget clamping, stream
truncation cost recovery
- [x] Cost validation on all 4 combos (direct×baseline, direct×thinking,
OR×baseline, OR×thinking) — `cost_usd` correctly persisted to
`PlatformCostLog` with `trackingType=cost_usd` (provider-reported, not
estimated)