Fix Agent Mode context-window reporting

[BetaFold3]

Fixes #506

Why

The context indicator could use three different denominators during one session:

• a static model maximum before provider activity,
• the internal 200K fallback while waiting for usage,
• and the provider-reported window after the first response.

Besides being confusing, this could produce incorrect percentages and silently discard valid usage readings above 200K.

This was especially visible with Codex GPT-5.5 and GPT-5.6: a fresh session could present the 200K internal fallback before runtime usage reported the actual window. GPT-5.5 was observed reporting approximately 258.4K after the first response. 353.4K in case of GPT 5.6.

What changed

• Resolve Claude’s configured auto-compact window from the effective launch environment and supported Claude settings files.
• Resolve Codex context configuration from user and project config.toml.
• Keep canonical model capacity separate from the effective reporting denominator.
• Show an unknown denominator as — instead of exposing the internal 200K fallback as fact.
• Carry configured Claude limits to the sidebar before the first usage event.
• Preserve spawn-authoritative values across asynchronous provisional resolution, model changes, and workspace changes.
• Retain valid large Claude usage readings while the canonical window is still unknown.

Bug found along the way

Claude result events may report usage for several models in the same session, including background Haiku work.

The translator previously selected the first positive modelUsage entry from an unordered dictionary. A background 200K model could therefore replace the main model’s 1M window, causing the sidebar to report 200K after the first response.

This change attributes usage to the tracked main model, handles normalized model variants deterministically, and resets attribution across live model switches and process respawns.

User-visible behavior

• A 1M-capable Claude model configured to compact at 400K reports 400K consistently, with the 1M model maximum shown separately.
• Codex GPT-5.5 and GPT-5.6 sessions show an unknown denominator before runtime usage, then switch to the provider-reported value when it arrives.
• Percentages and warning colors are suppressed while the denominator is unknown.
• Valid context readings above 200K are no longer dropped before the canonical window is known.

Validation

• Focused Claude/Codex resolver and context-usage suites
• make dev-provider-test FILTER=ClaudeSDKNDJSONTranslatorTests
• make dev-test FILTER=AgentRuntimeSidebarViewModelTests
• make dev-test FILTER=ClaudeContextUsageEstimatorTests
• make dev-swift-build PRODUCT=RepoPrompt
• make dev-lint
• Authoritative root/provider test lists and test-ledger reconciliation
• Manual CE debug-app checks for Claude 400K and Codex unknown-to-runtime-window transitions

[BetaFold3]

This PR does not modify the failing test or Codemap implementation. Mind helping with the re-trigger? @baron