fix: hide unknown Claude context usage

[lodystage]

Related issue

Internal same-repository fix; no issue is required for contribution intake.

Problem / pressure

Claude models whose runtime metadata omits a 1m marker temporarily appear to have a 200K context window, even when the SDK later reports a 1M authoritative window. The placeholder is presented as fact and creates a misleading first-turn experience.

Summary

• Advance the Claude ACP adapter to represent an unknown context window as null instead of 200K.
• Suppress numeric context usage updates until a cached, inferred, resumed, or result-reported window is available.
• Clear stale usage when switching to a model whose window is still unknown, then publish and cache the authoritative value from result.modelUsage.

Before / after

Before	After
A fresh Fable 5.1 session could display used / 200K until its first result.	Context usage stays hidden until the SDK reports the real 1M window.
Switching to an uncached model reused a guessed 200K window.	The old display is cleared while the new window is unknown.

Test plan

• pnpm --filter acp-extension-claude build
• pnpm --filter acp-extension-claude test:run — 813 passed, 20 skipped
• pnpm --filter acp-extension-claude check

Context handoff

Instructions for reviewing agents

• Review focus: Review nullable window seeding, every usage_update emission, and the model-switch clear path in the Claude adapter.
• Decisions to challenge: Confirm that the existing zero-sized ACP update is the right wire-level clear signal because ACP requires numeric fields.
• Plausible failures / evidence gaps: A backend that never returns result.modelUsage.contextWindow will keep usage hidden; no live authenticated Claude API run was performed.

Authoring context

• User goal / directives: Avoid presenting a guessed context size; hide context usage until a trustworthy window becomes available.
• Constraints / non-goals: Preserve sub-second session creation and model switching without reintroducing the blocking getContextUsage call.
• Risk-bearing decisions: Internal state uses null, while model switches emit numeric 0/0 solely to clear the previously persisted display.
• Destructive or irreversible behavior: No destructive behavior or migration; rollback is the previous Claude adapter submodule pointer.
• Deliberately not done or tested: No live authenticated provider request was run; deterministic adapter build, full tests, lint, and formatting cover the changed paths.
• Unknowns / confidence: High confidence for standard Claude SDK results; third-party backends that omit valid window metadata will intentionally show no context usage.