Streaming-Append Performance

The embedding contract's §5 requires append-mostly scenes (agent transcripts) to stay cheap. The Tier 0 spike measured this against 0.1.0 — it fails today, and the failure is quantified and scheduled.

The measurement

Transcript machine (Shared = Vec<String>, stable per-line node ids), 1,000 appends, per-turn set_shared + render_to_backend against MockBackend, 80×24 bounds:

Reproduce: cargo test --test embedding_harness -- --ignored --nocapture.

What's fine vs. what isn't

Fine: the committed patch surface per append is O(1) — one row insert plus a header update, bounded backend commands, no visible-window churn. Stable ids do their job.

Not fine: internal per-commit cost grows ~quadratically.

1. diff_render_ops is O(n²) — a linear scan of the previous op list for every op (and again for removals). Dominant at scale.
2. render_ops(bounds) re-runs full layout + lowering per turn — O(n).
3. set_shared / sample() clone the whole Shared — O(n) copy, no structural sharing.

The work item (0.1.x)

• HashMap-keyed O(n) diff_render_ops (id → index), the single biggest win
• Viewport-bounded projection so machines render the visible window only
• (Later) incremental Shared updates instead of whole-projection replace

At 1,000 lines, release-mode commits are sub-millisecond — the first Wallace slice is unblocked. At 10k lines the current path projects to ~80 ms/append, so the append-optimized diff lands before long transcripts ship.