fix(chat): stop the context-usage percentage flickering during streaming

[setkyar]

Fixes #110.

Problem

The context-usage capsule derives its snapshot from the last assistant message carrying a usage object. During a streamed turn a half-written assistant entry lands with an empty usage object, so contextTokens fell to 0 — the percentage jumped to 0% (or the capsule hid entirely) until the next reload put a real value back.

Changes

• collectContextUsage skips assistant messages whose usage totals zero when scanning backwards, staying on the last message that actually measured the context.
• updateContextUsage takes the snapshot it last rendered (previous) and returns the new one. A transient zero reuses the previous value, so once a value is on screen the capsule never flickers to 0% or hides. An empty model label likewise reuses the previous limit instead of falling back to the 128k default.
• createContextUsageController holds the first paint until /api/models settles — and paints anyway if that fetch fails — so the denominator does not visibly jump from the heuristic fallback to the registry value.

Tests

context-usage.test.js gains regression coverage: incremental streaming entries without usage don't move or hide the indicator; the limit survives a momentarily unknown model label; the first paint waits for the registry; and it still paints when the registry fetch rejects.

make check passes.

Known limitations

• The registry is pinned before the first paint, but the model label is not. If /api/models settles before worker status reports the model, the first paint uses the 128k default and adjusts once when the label arrives.
• Non-zero regressions are still possible if pi writes an intermediate assistant message with a small-but-positive usage. Monotonicity is deliberately not enforced — compaction legitimately drops the context.