fix(tui): derive context gauge fill from cumulative report deltas

[jkyberneees]

Problem

The header context-window gauge overflowed on any long multi-iteration session, e.g.:

● 100% 2.3M/1.0M · ∑ ⌂ 2.3M

while the backend context window was actually within budget.

Root cause

odek serve reports contextTokens — on both usage and done events — cumulative per run: the sum of prompt tokens across all LLM calls (agent.TotalInputTokens()). bodek treated it as the live window fill, so the gauge pinned at 100% once the cumulative total exceeded the model's context window. The ∑ ⌂ session total was correct all along.

Fix

Recover the live window fill client-side as the delta between consecutive cumulative reports — exactly the last request's prompt size, which drops again after odek trims history:

• new runCtxCum baseline on the model, updated on usage, consumed and reset on done (each run's cumulative restarts), and reset on session resume/clear
• winCtxTok is now only ever the per-request prompt size, never a cumulative figure
• comment corrections in internal/client and internal/tui/view.go documenting the real wire semantics

Tests

• New regression test TestGaugeDerivesFillFromCumulativeDeltas replaying the reported scenario (cumulative 600k → 1.4M → 2.3M against a 1.0M window): gauge shows 90%, and drops after a trim
• make fmt / vet / lint (0 issues) / test (full race suite) all green

Known limitation: mid-run the fill lags by one request's output tokens (the delta is prompt tokens only); it self-corrects on the next usage event. Exact per-request fill on the wire would require an odek-side change.