fix(ai,coding-agent): correct Copilot context window discovery and await it during startup

[spencerharmon]

Summary

Three bugs combined to make Copilot's 1M-context Anthropic variants effectively unusable: the model resolved to a smaller-window sibling and compaction fired well below the real ceiling.

Bugs fixed

1. Discovery host hardcoded. discoverCopilot hit api.individual.githubcopilot.com/models unconditionally, which returns HTTP 421 Misdirected Request for some accounts (business/enterprise tiers, and at least one individual account). Discovery silently no-op'd (if (!res.ok) return;) → 1M ids never merged into the registry → resolution fell back to a smaller-window sibling.

2. contextWindow used the wrong field. Discovery wrote max_context_window_tokens (total input+output+cache budget) to contextWindow. The compaction threshold (window − 16k reserve) should compare against the input budget (max_prompt_tokens). The threshold never tripped until the provider rejected the request.

3. Discovery races initial model resolution. discoverAnthropicCapabilities() was fire-and-forget at session startup. When the saved default model isn't in the static registry (true for any discovered-only id), resolution loses the race and falls back to a sibling with a smaller window.

Fix

• discoverCopilot tries the configured host first, then falls back through api.githubcopilot.com → .business → .enterprise. The host that responded is also used as the registered model's baseUrl so chat hits the same proxy that served /models.
• Both the capability cache and the registered model entry use max_prompt_tokens (fall back to max_context_window_tokens only when missing).
• createAgentSession and createAgentSessionServices await discovery before resolving the initial model only when the saved default isn't already in the static snapshot. Adds at most one /models round-trip to startup in that case; the fast path is preserved otherwise.

Verification

Direct simulation of shouldCompact against the patched code with a hypothetical 936k-prompt-cap model:

contextWindow: 936000  reserve: 16384  threshold: 919616
  tokens= 125529 → compact? false   ← previously a false-trigger
  tokens= 919000 → compact? false
  tokens= 920000 → compact? true    ← real ceiling

Live session confirmed: with this patch, resolution returns the discovered entry with the correct prompt-cap window from the first turn instead of compacting at the static-registry sibling's much smaller threshold.

Files

• packages/ai/src/providers/anthropic-discovery.ts — host fallback + max_prompt_tokens
• packages/coding-agent/src/core/sdk.ts — await discovery when saved default isn't in snapshot
• packages/coding-agent/src/core/agent-session-services.ts — same change for the services factory
• CHANGELOG entries in both packages

[spencerharmon]

CI failure here is unrelated to this PR — models.generated.ts is regenerated each CI run, and several test files reference model IDs that upstream providers no longer return (gpt-4o on github-copilot, qwen-3-235b-a22b-instruct-2507 on cerebras, google/gemini-2.0-flash-001 on openrouter). All 17 type errors are in test files this PR does not touch. Fix split out into #35.

[spencerharmon]

This CI failure is the same set of pre-existing failures present on main itself (run 27038697661 on d9a441e, the current main HEAD). The 9 failing suites — ws17.test.ts, package-manager.test.ts, settings-manager.test.ts, resource-loader.test.ts, etc. — all reference a .pi directory that no longer exists (the project uses .cave), and reproduce on a clean origin/main checkout with no changes from this PR:

$ git checkout origin/main
$ cd packages/coding-agent && npx vitest --run test/settings-manager.test.ts
# 3 failed | 15 passed

None of the failing tests exercise the files this PR touches (anthropic-discovery.ts, sdk.ts, agent-session-services.ts). Happy to file a separate cleanup PR for the stale .pi references if useful, but it's a much bigger surface area than the upstream-drift fix in #35.