Sprint 1: a 700 MB asset that survives a refresh
Sprint 1 (9 to 22 November 2026) had one measurable goal: a 700 MB model asset downloads in Chrome, survives a page refresh in the middle of the download, and is evicted cleanly when the storage quota is hit. This note is written from the four tester reports (T1 browser proofs, T2 Rust and interop gates, T3 release gate, S1 security review) and quotes no figure that is not in them.
Every number below was measured on an Apple-silicon development machine (18 cores, Google Chrome 157.0.8103.38 driven by Playwright 1.65.0, Node 22.22.3) and is a placeholder; the reference 8 GB Windows laptop without a GPU has not been measured yet.
What shipped
-
The five open contracts in
runtime/spec: Manifest, Capability profile, Event, Storage API and Backend adapter, each at 0.1.0 as JSON Schema 2020-12 plus TypeScript, with examples and five shared canonicalisation vectors. T2 compared all 26 vector files vendored into the control-plane crate with the runtime originals: all identical. -
The runtime asset manager: verified manifests (Ed25519 over a restricted
RFC 8785 canonical form), chunked Range downloads into OPFS or IndexedDB, resume from a
persisted part record, full SHA-256 verification before any byte is readable, and quota-driven
eviction. The runtime unit suite passed 278 of 278 cases in T1's run of
npm run ci. -
localdata: a set/get/delete/list/usage/clear SDK over OPFS, IndexedDB and an
in-memory driver. The browser behavioural suite ran the same case list on OPFS and IndexedDB,
values persisted across
page.reload()on both drivers, andquery()rejects withnot_implementednaming Sprint 4 (T1, 8 of 8 browser cases). -
pagetriage release preparation: formatting, clippy, 304 core unit tests plus
the golden, determinism and PDF suites, and the three-runtime determinism script (native,
Python wheel, WebAssembly) all green in T2's run; version stays 0.1.0 and no tag exists. The
release checklist in
docs/release.mdhas one row per definition-of-done item. -
glyphsift, tests first: 77 Devanagari cases in
tests/cases/devanagari.tomlcovering all 16 issue kinds, a harness wired tovalidate()and marked ignored with a message naming Sprint 2, and avalidate()that returns an explicit "not implemented" error in Rust, WebAssembly, the CLI (exit 3) and Python. Running the ignored harness fails on every case, as it should until the validator lands. -
cp-manifest and cp-sign (control-plane, private): the Rust half of the
manifest contract.
cp-sign signwith the RFC 8032 test seed over the unsigned example reproduced the signature inruntime/spec/examples/manifest.signed.example.jsonbyte for byte; the five canonical vectors and their signatures matched between TypeScript and Rust; all ten negative vectors were rejected by both with the named reasons; keygen writeskey.secretwith mode 0600 and refuses to overwrite (T2, 41 tests). - dockit: an installable PWA skeleton whose service worker hashes every precache entry before storing it, with a strict content security policy and no document processing. T1 found the content security policy free of remote origins and unsafe directives, and proved that a served copy with one precache hash flipped installs no worker and leaves no cache.
- This site: 30 papers in the tracker, each with a status and an area, a benchmarks scaffold whose every cell reads "not measured", and a build that refuses scripts, inline styles and third-party resources (T1, 9 of 9 build tests).
The 700 MB proofs
T1 ran the runtime demo page against the repository's synthetic range server
(test/browser/range-server.mjs), which generates the asset bytes from a seeded
xorshift64* stream so no large file is stored anywhere. A separate Node script
(test/tools/asset-digest.mjs big 734003200) computed the expected SHA-256 of the same
stream independently of the browser code:
0f0664c252212d3be1f6b783b42b8e3507ab9b218984dd25b52a2be561a9db35, in 0.85 s.
| Scenario | Store | Result | Measurement (Apple-silicon placeholder) |
|---|---|---|---|
| Full download of 734,003,200 bytes, SHA-256 verified against the independent digest | OPFS | pass | 4060 ms wall, 180.8 MB/s, peak JS heap 2,793,500 bytes |
| Same | IndexedDB | pass | 2561 ms wall, 286.6 MB/s, peak JS heap 2,795,400 bytes |
Reload the page mid-download, then call ensure() again |
OPFS | pass |
276,824,064 bytes written at reload (37.7%); exactly one request after the reload,
Range: bytes=276824064- with If-Range carrying the server's
strong ETag, answered 206 with 457,179,136 bytes, which is the size minus the start; final
hash matches; no part file left
|
| Same | IndexedDB | pass |
285,212,672 bytes written at reload (38.9%); one request
Range: bytes=285212672-, 206, 448,790,528 bytes sent; resume leg 2805 ms
including the reload
|
Server with --corrupt-hash |
OPFS | pass |
ensure() rejects with hash_mismatch, asset.verify.failed
emitted, no index row, no committed file, no part file; getBlob() rejects with
not_verified
|
Server with --ignore-range on a resume |
OPFS | pass |
one request answered 200 with all 734,003,200 bytes; asset.download.restarted
with reason status-200; final hash verified
|
| Committed file truncated by one byte | OPFS | pass |
getBlob() rejects with not_verified, the index row is removed,
asset.verify.failed reason size_changed
|
| Manifest entry pointing at another origin | both | pass |
blocked with policy and asset.download.blocked before any request;
the server log for it is empty
|
| Network egress across every scenario | both | pass |
the union of the page's requests and the server log contains only 127.0.0.1 URLs: the demo
page's own files, /manifest.json and /assets/*
|
Peak JS heap stays near 2.8 MB because the manager streams 8 MiB chunks straight into the store;
the figure does not include the OPFS or IndexedDB buffers held by the browser process. The
capability probe (detectCapabilities()) resolved well inside the 2 s budget (147 ms
test wall time including the page round trip) and its details object carries only the
allow-listed keys.
What did not pass in round one
The third clause of the sprint goal, "evicts cleanly when the storage quota is hit", failed in T1's first run and is the reason the sprint was not closed on the first pass.
-
OPFS commit at the quota edge. With the origin quota set to 1.5 GiB through
Chrome DevTools Protocol (
Storage.overrideQuotaForOrigin), T1 downloaded A (700 MB) then B (400 MB), read A, then downloaded C (500 MB). The manager evicted B with reasonquotaas required, but C's commit then failed withAssetError [store] commit failed: concatenating the chunk files into the final file needed a transient second copy of the asset that the eviction had not reserved. Left behind were an orphan unverified file for C and 65 part files;reconcile()removed the orphan but not the part, and every laterensure(C)received 416 from the server and failed at commit again. The fix round assigned this to the runtime lane as the sprint blocker: bound the transient to one chunk, reserve it, evict and retry onQuotaExceededError, leave no orphan or stuck 416 state, and route every store write through one error-mapping helper. T1's retest of the same scenario (OPFS, 1.5 GiB CDP origin quota, A 700 MB then B 400 MB,getBlob(A), then C 500 MB) evicted exactly B with reasonquota(419,430,400 bytes), left A and C verified, listed exactly A and C inindex.json, left no.partfile and no orphan, and took 9582 ms wall for the whole sequence on the Apple-silicon development machine (placeholder figure; evidenceT1/report.md, fix-confirmation row T1-B1, andT1/retest1/t1-retest1-quota-opfs.json). The commit-at-the-quota-edge blocker is closed for OPFS. -
IndexedDB quota behaviour was not provable with the CDP override. Chrome 157
accounted 1,053,040,640 bytes of IndexedDB usage for 1,677,721,600 bytes of stored chunks and
never threw
QuotaExceededErrorunder the 1.5 GiB override, whilenavigator.storage.estimate().quotaignored the override entirely, so neither the reactive nor the estimate-based eviction path fired and A, B and C all stayed stored. The fix round switched the IndexedDB eviction proof toquotaBytesinjection, and the IndexedDB half did not evict in that retest either: with the 1.5 GiBquotaBytesoverride active,navigator.storage.estimate().usagereported 723,398,656 bytes for 1,153,433,600 bytes of stored chunks after A and B (ratio 0.627), so the manager's estimate-based accounting projected that C would fit, emitted noasset.quotaand noasset.evict, and A, B and C all ended verified (T1/retest1/t1-retest1-quota-indexeddb.json). This is bug T1R-01, open against the runtime lane at the time of writing: the asset manager needs its own byte accounting for the IndexedDB store rather than trusting Chrome's estimate. T1's round-2 result is not yet in, and this write-up does not claim the IndexedDB half closed. Real-quota IndexedDB eviction and disk-pressure behaviour remain deferred to the Windows laptop. -
The runtime browser suite as committed did not run. 9 of 12 Playwright cases
failed because the demo page called the asynchronous
list()synchronously, the IndexedDB part record was read from the wrong store, and the progress throttle of 250 ms skipped the 38 to 42 percent reload window. T1 produced the proofs above on a copy of the demo with a one-line fix; the fix round repairs the demo and helpers and adds them to the type check. -
Smaller items. localdata's
npm run cifailed on vendored-contract drift (one union type re-wrapped by a formatter; the consumer-side fix is to re-run the sync script before the contracts are committed); one runtime unit test about abort-then-resume failed in T2's run; the glyphsift Python wheel did not build because itspyproject.tomlreferenced files absent from the crate; dockit's service worker served an unverified network response when the runtime re-verification of a precached file failed after a cache miss (it now answers with the offline page or a 503); and dockit had no repository test for the tampered-precache behaviour T1 had proved by hand (it now has one).
Security review
S1 found no Critical or High issue, one Medium (the manifest schema's algorithm enumeration
leaves no room for agility) and twelve Low items, all assigned to the fix round, and wrote the
threat model, findings register and pre-pilot checklist under
runtime/docs/security/. The accepted risks are on the record with target sprints: no
key revocation, no transparency log, no Sigstore compatibility, a single-signature threshold, trust
in the browser clock for expiresAt, same-origin co-tenant trust for OPFS, and the
unmeasured IndexedDB fallback on the 8 GB laptop. T3's release gate found every CI action pinned
by full commit SHA and matching the published tag refs, per-job permissions everywhere, no
pull_request_target, no secrets, and zero hits in the clean-room sweep across all
seven repositories.
pagetriage v1.0.0 was not tagged
The plan asked for pagetriage to be "release-ready for the v1.0 tag" this sprint. It is not, and
nobody tagged it. docs/release.md lists the seven definition-of-done items; T3
verified every row against the repository and found only row 3 (README limitations section)
Done. Rows 1 (tests first and CI green on Linux, macOS and Windows: the Windows job has never
run because nothing has been pushed to the organisation yet), 2 (benchmark on the 8 GB laptop),
4 (signed artefacts and manifest: no key exists), 5 (changelog and semver tag), 6 (installable
from a clean machine: nothing is on crates.io, PyPI or npm) and 7 (write-up) are Open. The
version stays 0.1.0, the 1.0.0 notes live under Unreleased in the changelog, and the owner has not
been handed a tag to approve. T3's recommendation was either to close rows 1, 2, 4, 6 and 7 and
then tag 1.0.0, or to tag an explicitly pre-release 0.9.0 whose notes name the Open rows.
What is next
Sprint 2 brings the glyphsift Devanagari validator and its corruption benchmark, the first real backend adapters behind the backend-adapter contract, and the deferred runtime hardening (cross-tab orphan race, IndexedDB file-size trust, the acquire refcount gap). The 8 GB Windows laptop measurements, the real-quota IndexedDB eviction behaviour and every Windows CI run wait for the owner's first push and the laptop itself; until then every figure on this site stays labelled as a placeholder.
Evidence for every statement above is in the tester reports and their artefacts (download and resume JSON reports, Playwright traces, server request logs, gate logs). Prose on this page is CC BY 4.0; attribute "Airplane Mode Labs" with a link.