fix(runtime): persist contextRemaining on the stored TokenUsageMessage so TUI ctx survives transcript rebuilds

[me2seeks]

Problem

The TUI status line permanently degrades to ctx ?/<window> (#3371 fallback) even for providers that report per-step usage, because the stored TokenUsageMessage never carries contextRemaining.

In packages/runtime/src/ai-sdk-backend.ts the turn-end usage write has two asymmetric paths:

• The live TokenUsageEvent pushed to the session queue includes contextRemaining (added by fix(runtime): emit contextRemaining in token_usage events #1072).
• The durable TokenUsageMessage written via appendMessage(tu) does not — the field is absent from the object literal.

The TUI rebuilds usage from stored messages in replaceTranscriptWithStoredMessages (packages/cli/src/pi-transcript.ts), so after any transcript reload (abort refresh, session resume, reconnect) contextRemaining is washed back to undefined and the status line falls back to ctx ?/<window> — even though the runtime event ledger (runtime_events.actions.tokenUsage) proves the value was computed (e.g. contextRemaining: 230068 for a 262144-window model).

cacheHit/cacheMiss survive the same rebuild (they are persisted), which is why cache NN% stays correct while ctx reverts to ?.

Expected

The persisted TokenUsageMessage carries contextRemaining (same conditional-spread value as the live event), so the ctx segment survives transcript rebuilds.

Scope

• Move contextRemainingForUsage computation above the stored tu literal and add the field to it (schema already allows it: packages/core/src/usage-record-schema.ts, packages/core/src/session.ts; replay/read-model paths already preserve it: runtime-event-backfill.ts, runtime-event-read-model.ts).
• Extend the ai-sdk-backend usage tests to assert the stored message includes contextRemaining.

Refs: #1067 (origin), #3371 (degrade-to-? fallback), #1422 (preserve token usage fields across replay).

[me2seeks]

I'm working on this.

[me2seeks]

Field report consistent with this issue (posted via AI agent on behalf of the reporter):

• Surface: TUI, model k3-256k (context window 262,144), thinking:high
• Session shape: single turn ran >1h10m with many Bash tool calls (local CI run + polling); user steering messages and interrupts occurred during the turn, so a transcript rebuild (abort refresh) likely happened mid-session
• Symptom 1: status line showed ctx ?/262k (feat(cli): status line should degrade gracefully when ctx/token usage is unavailable #3371 degrade) for the whole observed session — no usage percentage was ever visible
• Symptom 2: the turn then ended with Error: Turn failed: context_budget_exhausted with no prior visible warning (the >80% yellow / >95% red cues from fix(cli): statusline density, color fallback, and activity-strip spacing #1064 never had a chance to show)

If this provider does report per-step usage, this session is a real-world repro of the wash-out described here: after a transcript rebuild the durable TokenUsageMessage lacks contextRemaining, the status line falls back to ?/<window>, and the user loses exactly the visibility that would have warned them before the budget verdict. (If the provider does not report per-step usage, the ? is the expected #3371 degrade and orthogonal to this fix — the reporter believes usage is reported.)