- /api/overpass: a loopback socket with forwarding headers (a local reverse
proxy) is admitted from the remote pool under a socket-keyed proxied: bucket,
never the owner's direct-loopback window; the upstream limiter uses the same
key. Regression: owner cache hits survive an exhausted proxy bucket.
- Arrival check: a reveal timeout keeps the settle angle it measured; the new
observeArrivalRotationWithRetry retries only inconclusive timeouts, so a
real violation cannot be retried into a pass. Regression: wrong settle angle
-> reveal timeout -> clean second attempt still fails.
- SECURITY.md, README, changelog and CURRENT-STATE: scope the guarantee to
admission isolation, name the forwarding headers a local proxy must send,
and state that the upstream budget stays shared.
A headless SwiftShader run showed a 926 ms first frame on the military probe,
which no wall-clock budget can accept: the 1 s idle refresh may already have
run inside it. Raise moveEnd and wait one frame before each attempt so the
settle frame is measured against a pass that just ran, and treat a timeout
(never a wrong angle) as inconclusive with one more attempt. Document why the
per-phase budget stays at half the refresh interval.
Co-Authored-By: Claude Fable 5.1 <[email protected]>
Address gekh's review of 69a87fe2:
- P1: /api/overpass now uses admitSameSiteKeyless, like /api/terrain/heights,
so an instance behind the README's reverse proxy keeps its Overpass-backed
layers; cross-site shapes are still refused and forwarding headers no
longer launder them. SECURITY.md, the gate header and CURRENT-STATE say so.
- P2: admission keys IPv6 peers by /64 and folds IPv4-mapped peers
(clientNetworkKey), and loopback has its own 300/min window outside the
900/min global, so a few LAN peers cannot lock the owner's browser out.
- P2 (track): the settle and reveal frames each get their own 500 ms budget,
measured from moveEnd and from the settle frame respectively, so a ~300 ms
SwiftShader frame no longer trips the reveal deadline while both stay
inside the 1 s idle refresh.
- Smaller: drop the dead settings-block check and stale probe comment in
query.js, clamp whitespace-tolerant [ timeout : N ], and add tests for the
global window, the /64 keying, loopback headroom, proxied admission and
cache hits spending admission but not upstream budget.
Co-Authored-By: Claude Fable 5.1 <[email protected]>
- taiwan.js: a failed download serves the last good table (up to 24 h)
and is not retried for an hour. The table host does not answer from
some networks outside Taiwan, and each attempt held the shared catalog
refresh for the ~10 s connect timeout.
- Own server/providers/cctv/taiwan.js in package-boundaries.json
(check:boundaries failed).
- Decode numeric character references; name falls back to the CCTVID
when a row has no road details.
- Credit names the dataset and links the OGDL Taiwan licence.
- Document CCTV_TAIWAN_URL; list Taiwan in the CURRENT-STATE source table.
- Tests: isolate catalog env, reset loader state, cover the byte cap,
failed fetch, backoff, last-good fallback, cap clamping, redirect with
a valid body, name fallback, numeric entities and UUID camera ids.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Fetch keyless DATA.GOV.HK camera locations at runtime and pin stills to
tdcctv.data.one.gov.hk with required attribution. Default cap 1,100 covers
the ~1,013-camera TD catalog, prioritized across Hong Kong Island, Kowloon
and the New Territories.
Co-authored-by: Cursor <[email protected]>
- Google 3D: the OPEN button re-measures a ground whose tiles are still
loading (a few times, a second apart) and on a map change, not only
after the next camera move.
- Mapillary: two mounts on one host during the first download share one
build instead of leaking a second WebGL viewer.
- Retired floating-panel keys (v8 panelPos, panelFloatHintShown) are
dropped from localStorage; dead inline-position resets, the drag JSDoc
and the dead qa-ui-disposal observer branch are gone.
- Docs: z ladder, panel chrome, position storage and the fixture gate's
actual coverage; changelog fragment for #974 (old .gev-cache/mapillary
can be deleted).
- ru/es: the layer is named as everywhere else in the UI.
- Fixtures and test fakes: unused exports removed; a fake credit removal
of an absent credit no longer drops another one; the tile test reset
frees upstream slots.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
page.waitForFunction polls on requestAnimationFrame by default. In a
minimised or hidden headful window rAF stops while timers keep running, so
each 120 s slice only checked the predicate when it started: a drain that
finished in the last slice was recorded as a FAIL on the sign-off path.
- waitForPredicate replaces waitForFunctionSliced and the other three
waitForFunction calls: one short page.evaluate per tick (1 s for the
drain, 250 ms elsewhere), a monotonic performance.now() deadline, retry
across a context swap, and a throwing predicate fails the run instead of
silently spending the budget.
- One in-page window.__qaRenderFrames(count, capMs) helper serves both the
gizmo-drag and the sibling-owner frame waits.
- The sibling pick counts as a stalled renderer only at 0 of 3 frames, so a
slow SwiftShader frame still reports a real miss; the record flattens and
prints the frame count once.
- Review screenshots log and continue when a capture fails.
- Launch with pipe: true (no localhost CDP port) and protocolTimeout
300_000 like the other harnesses; drop the occlusion flag that puppeteer
already passes by default.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
When the sibling-owner pick misses after fewer than 3 rendered frames, the
cause is a stalled renderer, not a pick regression. Record it as a failure
on the headful sign-off path and as inconclusive under headless GL, with
"renderer stalled: N/3 frames" in the detail. A missed pick after 3 frames
still fails in both modes. Also corrects the catalog-size comment (~3,700).
Co-Authored-By: Claude Opus 5.5 <[email protected]>
The headful drain budget is uncapped (camCount * 800 + 30000 ms, ~50 min
at N≈3,700), but each page.waitForFunction is one CDP call that puppeteer
aborts after its 180 s default protocolTimeout. The .catch then recorded
"timed out waiting for drain on the real-GPU sign-off path" after ~3 min.
- Poll the drain in <=120 s slices (waitForFunctionSliced) so no single
CDP call outlives protocolTimeout, whatever the catalog size.
- Launch with protocolTimeout 240 s, headroom above the largest slice and
the 60 s fixed waits.
- Bound the sibling-owner 3-frame postRender wait at 10 s and report the
frame count, so a renderer that stops producing frames fails with
evidence instead of hanging the evaluate.
- List --disable-backgrounding-occluded-windows explicitly. Puppeteer
25.10 already passes it by default, so this documents the headful
contract rather than changing behaviour.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Conflicts: the Street Level and Gemini voice strings both appended to the
es/ru catalogs, and both sides' modules to package-boundaries.json — kept
both. scripts/qa-panel-resize.mjs stays deleted (floating panels removed).
Co-Authored-By: Claude Opus 5.5 <[email protected]>
On CI's SwiftShader the ground-clamped cones take a few frames after a
sequence is selected; a single pick search found none.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
- A background catalog refresh that rejects now keeps the old catalog for
another TTL, as an empty refresh already did, so a refresh that keeps
failing is retried once per TTL instead of on every request.
- Trim notes are keyed by the source list they describe, and /sources asks
for the list it served, so a request answered during a refresh never pairs
the old cameras with the new list's notes.
- Tests: a failed background refresh (one warning, no unhandled rejection,
no retry before the TTL, trim notes stay with their list) and an empty
background refresh (old catalog kept, no upstream fetch before the TTL).
The race assertion's message now names the failure; its timer is cleared.
- The Tarktee test no longer imports the new constant, so on main it fails
on its own assertion instead of taking the whole file down.
- precompute-cctv-heights waits 60 s for a cold catalog, past Tarktee's 30 s.
- CURRENT-STATE notes the 30 s bound and the retry spacing.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
refreshCctvFocusStyles passes cameraColorFor as baseColorFor, but only
refreshCoverageStyles was tested. The new rendering.test.mjs runs the
focus pass through createRendering with stubbed focus services and
checks that an idle video camera fades from violet, a still from cyan
and the active camera from amber. It fails if the focus callback goes
back to the old active/idle check. Also adds the missing semicolon on
the policy.js import in cctv.test.mjs.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Said Turksever