fix(app): scale MAX_CONTEXT_SIZE for 1M models + render alwaysShow before usage streams in

[g0rdonL]

Summary

Two related fixes to the Always Show Context Size indicator in AgentInput. Originally only #910; on rebase I traced #1274 to the same render path and added a one-line gate fix that pairs naturally with the model plumbing.

Commit 1 — Closes #910 (scale MAX_CONTEXT_SIZE for 1M-context models)

• Context-remaining indicator was hardcoded to 190K, so any session on a 1M-context Anthropic model (e.g. claude-opus-4-7[1m]) clamped to 0% within minutes and the toggle appeared broken.
• Plumb the assistant model field from the raw Claude API event through NormalizedMessage → reducer latestUsage → Session.latestUsage → AgentInput.usageData.
• AgentInput now picks the limit dynamically via getMaxContextSize(model): 950K for model ids matching [1m], 190K otherwise.

Commit 2 — Closes #1274 (render alwaysShow before usage streams in)

• The outer gate at the call site short-circuited to null whenever usageData.contextSize was 0 or undefined:

  const contextWarning = props.usageData?.contextSize
      ? getContextWarning(...)
      : null;

  …which swallowed the alwaysShow signal that getContextWarning itself honors (line 311). Two failure modes this addresses on Android (and silently on iOS/web — same render path):
  1. Fresh session with no agent reply yet — latestUsage is undefined, indicator never showed.
  2. Sessions where the reducer resets contextSize to 0 on session events (reducer.ts:324, 338) with no subsequent assistant message carrying usage.
• Gate is now alwaysShow || contextSize, and getContextWarning receives contextSize ?? 0. On fresh always-show sessions the indicator renders "100% remaining" until real usage arrives.

Files changed

• sources/sync/typesRaw.ts — capture message.model when normalizing assistant messages; add model?: string to NormalizedMessage.
• sources/sync/reducer/reducer.ts — thread model through processUsageData / ReducerState.latestUsage / ReducerResult.usage.
• sources/sync/storageTypes.ts — add model?: string to persisted Session.latestUsage.
• sources/-session/SessionView.tsx — thread model into the memoized usageData source mapping.
• sources/components/AgentInput.tsx — getMaxContextSize(model) helper + rewrite the contextWarning gate to honor alwaysShow when usage is missing.

Test plan

• pnpm typecheck in packages/happy-app passes.
• pnpm test --run in packages/happy-app — identical failure baseline to main (11 failed / 466 passed / 57 skipped on both branches; failures are pre-existing markdown, GitHub API, settings-defaults, and model-mode tests untouched by this PR).
• Manual: open a session on claude-opus-4-7[1m] with Always Show Context Size enabled; the percentage should reflect 1M-token usage rather than sticking at 0%.
• Manual: open a 200K-context session (Sonnet/Haiku); behavior unchanged.
• Manual: enable Always Show Context Size on a fresh session (no agent reply yet) — the indicator should render "100% remaining" immediately rather than waiting for usage.

Notes

• [1m] detection is /\[1m\]/i against the model id returned by the Claude API stream. Happy to switch to an explicit lookup table if you'd prefer.
• model is optional throughout — pre-existing sessions and non-Claude flavors keep working with the 190K default.
• Rebased onto main (270+ commits since the original PR); SessionView.tsx was reconciled with the new memoized usageData introduced in main.

Closes #910
Closes #1274