[sergeytimoshin]

The browser prover gets the proof timing report the mobile SDK has: four scenarios for every shape, before and after a change, commented on every pull request, with a history of main.

• examples/browser/scripts/proof-timing.mjs: drives the benchmark page in headless Chrome (selenium, as test:browser) through first run including the key's fetch and decode, warm, first run after a page reload (raw-key cache) and repeated, for transfers of 1 to 24 notes and a merge of 24 (ZOLANA_TIMING_SHAPES narrows it); a failed or stalled shape prints an error line and the others run in a new browser.
• scripts/proof-timing-report.mjs (+ test): the table with before/after rows, the download bracket, the stage split and failed shapes; proof-timing-ci.sh takes the baseline from the history and writes the job summary and one PR comment; proof-timing-history.{mjs,sh} append each main report to the proof-timing branch and rewrite its README.
• wasm/prover-wasm/testdata: transfer-{1,2,4,8,16,24}x2.json and merge-24x1.json, valid requests on trees of their own (the mobile SDK's tools/prove-fixtures wrote them), with Go tests of their shapes; build-wasm.sh stages the transfer requests.
• ci.yml: a bench job beside check, cancelled by a newer push to the PR, and a proof-timing-history job with contents: write after each push to main.
• README and CHANGELOG.

On this branch's runtime (0.1.0) 24x2 crashes the browser and the merge stalls at the kernel's 1 GiB memory cap, so the report shows both failed until #21 and #22 bring runtime 0.2.0. The proof-timing branch appears with the first push to main after merge; until then the PR comment has no before rows.

[github-actions]

Proof timing: chrome

After: this pull request at 049c300. No report of main to compare with yet.

Medians of one browser, by shape (notes → outputs). Runner numbers are for comparing commits, not for how fast a device is.

scenario	1 → 2	2 → 2	4 → 2	8 → 2	16 → 2	24 → 2	merge 24 → 1
first run, including download	12.17 s (download: 2.29 s, 8.8 MiB)	15.89 s (download: 0.62 s, 15.6 MiB)	30.89 s (download: 2.69 s, 28.9 MiB)	58.51 s (download: 3.81 s, 55.9 MiB)	112.74 s (download: 5.85 s, 109.5 MiB)	failed	failed
warm	1.64 s	2.37 s	4.29 s	8.25 s	16.03 s	failed	failed
first run after restart	2.19 s	3.21 s	5.45 s	10.03 s	19.26 s	failed	failed
repeated after restart	1.39 s	2.37 s	4.32 s	8.20 s	16.02 s	failed	failed

stage	1 → 2	2 → 2	4 → 2	8 → 2	16 → 2	24 → 2	merge 24 → 1
key load	0.25 s	0.42 s	0.70 s	1.33 s	2.57 s
witness	0.00 s	0.00 s	0.00 s	0.00 s	0.00 s
prove	1.48 s	2.31 s	4.23 s	8.16 s	15.94 s
verify	0.02 s	0.01 s	0.01 s	0.02 s	0.03 s

Failed:

• 24 → 2, fresh run: Wait timed out after 300008ms
• merge 24 → 1, fresh run: Wait timed out after 300000ms

[sergeytimoshin]

Superseded: this change reached main as cc95006 within #23 (together with runtime 0.2.0, which is why the bench here was red: on this branch's base, runtime 0.1.0's 1 GiB kernel cap stalls 24→2 and the merge). The bench is green on main.