perf(tui): single-pass token accounting on per-turn pressure paths

[AdityaVG13]

perf(tui): single-pass token accounting on per-turn pressure paths

The per-turn metadata build and the compaction decision each re-walked the full transcript more than once per step, and the streaming render re-processed the whole accumulated message on every chunk. This series makes those walks single-pass with byte-identical behavior, pins the equivalence with tests, and fixes one cache-invalidation bug found along the way.

Path	Before	After
Compaction decision per step under pressure	2 full-transcript estimate walks	at most 1
<turn_meta> build per user or steer message	full-history deep clone, then estimate	estimate in place, arithmetic delta
Billed input above threshold	full estimate walk	skipped, guards preserved
Session restore	stale estimate could serve capacity checkpoints	cache invalidated, recomputed
Streaming render per chunk, no math present	full-content copy and transform	one byte scan, zero copies

One estimate per compaction decision

compaction_decision_with_billed called the same pure estimator twice per step once pressure existed: once in the pressure gate and once in the prune projection. It is now computed at most once and threaded to both consumers. Guard order (gate, prune projection, too-few-messages, retained-floor) is untouched and decisions are identical for all inputs. When provider-billed input alone proves pressure, the estimate walk is skipped entirely without bypassing the guards that keep a compaction loop from re-firing on every step.

No full-transcript clone in the per-turn metadata build

The builder deep-cloned the entire message history just to append the composer text before estimating. It now estimates the history in place and adds the one-message delta arithmetically. The subtle part: the estimator inflates the summed byte count by a factor of 1.5 rounded up as a whole, so the delta is floor(own * 3 / 2) plus one exactly when the running byte sum is even and the composer text contributes an odd count. This is proven against a clone-and-estimate reference over 80,000 parity combinations and pinned per case by a test that also covers replayed reasoning, the arm where a naive helper undercounts.

Session restore invalidates the estimate cache

Session sync assigned the restored message list directly to the session field, bypassing the revision bump the token-estimate cache keys on, so a restored conversation could be priced from a stale estimate. The bump is restored and a test pins that the cache recomputes after a restore.

Zero-copy LaTeX pre-pass

The streaming render ran the math transform over the full accumulated content on every chunk, before the incremental markdown cache that itself only renders deltas. The transform now returns a Cow<str>: one byte scan for the three opening delimiters decides between borrowing the input untouched, the overwhelming case for streamed prose, and running the transform. Output is byte-identical either way. A test pins the borrowed path and that the \( form is not missed by the fast scan.

Testing

cargo fmt clean. compaction:: and latex_render::: 62 passed. Engine preview, pressure, and restore tests: 14 passed. cargo check --all-targets: clean, no warnings.

[Hmbown]

Thank you, Aditya. I verified both patches in this PR are already on main through #5667 (6ef5f87d8182eb38a18cfff99e58bbf7ef01b20c). The original commits and landed commits have identical stable patch IDs, and the landed commits preserve AdityaVG13 <adityavgcode@gmail.com> as author.

Landed commits: 692cf1efb and 5edb8c3f9.

I’m closing this duplicate PR because the contribution is already shipped on main; merging this stale head again would duplicate the patches.