The served Vibe preview was eight items deep, and every one of them cost
a track fetch on each advance while buying nothing but a longer Up next
list. Three is enough to show where the stream is going.
The audio prefetch was the opposite problem: it only began twenty
seconds before the end, so a phone that lost signal in that window
arrived at the handover with nothing buffered. It now starts fifteen
seconds into the current track, which gives the rest of the song to pull
the next one down. Buffering that early means a replan can change the
answer, so the idle element is re-pointed when it does.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
A locked screen drops the radio, changes network or dozes, and one
request fails. Both halves of an advance treated that as the session's
fault. The outbox kept the failed event at its head with the comment
that a later retry would pick it up, but nothing ever triggered one, so
it sat there while the caller was rejected. advanceVibe then cleared the
prefetched future and paused, destroying a plan that was still valid.
Transient failures now hold the outbox entry unsettled and resend the
same event id on a backoff, so no duplicate feedback reaches the
director. The retry around the serve sits on the serve alone: serving a
version is idempotent, while replaying the whole advance would report a
second outcome for a track heard once. Each wait ends early when the
browser says the network is back, which is the moment that matters when
a screen unlocks. A drop that outlives every retry leaves the future
intact to carry on from.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The owner reports its position every ten seconds, and the server
publishes a snapshot to every device on anything that touches the
session. The owner was writing those snapshots back into its own store,
so each one dragged playback back to its last report. Reloading another
tab did it too, since registering a device publishes a snapshot: the
phone jumped back to whatever position was on record, which right after
a track change is zero, and the song started over.
A device that already owns the session now ignores incoming snapshots
and stays the authority on its own position. Snapshots still apply when
ownership moves to or away from this device, which is what that branch
was written for. Moving audio between devices also flushes the real
position first, rather than handing over a ten-second-old one.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Adds the manifest, icons and service worker that make the app
installable, and offers it as a toast once Chrome says it qualifies.
Declining snoozes the offer for a month; installing ends it.
A waiting service worker never activates on its own. Reloading the page
under a listener to swap in a new build would cut the song they are in
the middle of, so updates land on the next cold start instead. Audio is
kept out of the cache entirely: range requests and multi-megabyte bodies
do not belong in a shell cache. Artwork is cached, and the SPA
navigation fallback denies /api so it cannot swallow the event stream.
Installed on Android the app paints edge to edge, so the transport pads
itself past the gesture bar. MediaSession gains setPositionState, which
is what gives the notification shade a seek bar that moves.
Three things kept the bundle from ever being compressed, each hiding the
next: the nginx image ships with gzip off, gzip_proxied defaults to off
and skips anything carrying a Via header, and gzip_http_version defaults
to 1.1 while the host proxy speaks 1.0. With those fixed and the pages
split per route, the first load goes from 555KB to 60KB of app code plus
a vendor chunk that survives redeploys.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The v1 `docker-compose` binary on the deploy host fails with
KeyError: 'ContainerConfig' when it recreates a container, so every
instruction pointed at a command that cannot finish a deploy.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The queue panel drove the playback store directly, so its buttons played
locally while another device held the audio. Both control sets now go
through one transport that forwards a press when the audio is elsewhere.
The panel's artwork is capped against viewport height too: at full width
on a phone the square alone pushed Up Next off the screen.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
One device holds the audio; the rest watch the same session over an
event stream and act as remotes. Picking a device hands the audio over
at the position the previous one reported, and that device stops.
Also centre the command palette with margins instead of a translate:
animate-rise sets its own transform and dropped the offset on mobile.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Acquisition ran yt-dlp without --embed-metadata, so every download
arrived untagged. The scanner then stored the video id as the title and
"Unknown Artist" as the artist, the vetted-candidate tag check rejected
the mismatch, and all 18 acquired tracks were hidden and retired.
- Pass --embed-metadata so downloads carry real tags.
- Let a scan take fallback title/artist from the candidate, for sources
that still ship untagged files.
- Install Deno alongside yt-dlp: YouTube guards some formats with a JS
challenge yt-dlp must execute, and no other runtime is enabled.
- Dedupe candidates by artist and title. The (source, external_id) key
misses the same song reaching us under two Deezer release ids.
Also carries the in-flight discovery work this builds on: the
Recommendations page replacing Discover, the discovery source service,
and the acquisition spec tests.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The image proxy allowlist covered none of the hosts enrichment writes, so every
album cover answered 403: 305 covers sit on coverartarchive.org (which 302s to
archive.org and on to an ia*.us.archive.org node), 126 on mzstatic, 115 on
dzcdn, 26 on discogs.
Artist images were empty for a different reason — no working source. Fanart
needs a key the worker does not have, TheAudioDB and Discogs 404 on most names,
Wikidata needs an MBID that 638 of 734 artists lack, and Last.fm stopped
serving photos. Deezer needs no auth and its host is already allowlisted.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
A pass over the whole app against the Ethos laws, then a focused pass on
Vibe with the operator reviewing each change.
Across the app:
- The player bar restores the last track it played, paused at zero, so a
fresh tab opens on where the listener was instead of "nothing playing".
- Track titles link to their album, matching the artist links beside them.
Playback stays on the artwork tile; a title that played was the surprise.
- The search field is bg-bg2. Tailwind cannot alpha-modify these var()
colors, so bg-surface0/70 emitted no rule at all and the input fell back
to the UA's white.
- Row hover is light falling off to the right, not a flat slab.
- The artwork placeholder can drop its note glyph, so TrackRow no longer
layers a play icon on top of one.
Vibe:
- A seeded Vibe plays its seed first. The seed sits in front of the durable
plan without being part of it, so the first advance consumes it locally
and reports no plan feedback.
- The queue drops a second recording of a song it already holds — same
title, different track id, which id-based dedup let through.
- Up next is read from the queue rather than the plan preview, since the
seed is not a plan item.
- The header carries the live profile (energy, discovery, goal) and both
verbs. Keep is gone: letting a track finish already reports `completed`,
which the director weighs the same.
- The aura is one warm diffuse blob in the page background, warm-hued only
and quieter on mobile.
- Compact artwork is 32px. It was h-8 w-8, which this remapped spacing
scale renders as 64px inside a 44px row, and that overflow was the
"stacked" look.
Verified by render at 1440x900 and 390x844, no horizontal overflow at
either. 26 frontend and 122 backend tests pass.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The director had no idea what hour or season a session started in, so a
22:00 weeknight and a Sunday morning drew from the same pool. The client
now sends localHour, weekday, month and an optional timeZone; the route
validates and bounds all four, the coordinator threads them through, and
the generators use them as scoring signals.
Also exposes GET /library/stats, which the Home page counts read.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The isPlaying subscriber is unselected. Every store write during Vibe's
feedback/replan handshake called play() on the element that had just ended.
That replayed its final buffered milliseconds until the next source loaded.
Gate that subscriber and the seek subscriber on an actual value change, and
never resume a finished element.
Then close the gap the handshake leaves behind. The engine now drives two
<audio> elements. The next track buffers into the idle one 20s early. The
handover starts before `ended`, so the round-trip happens under the outgoing
tail. With a crossfade, that tail fades out under the new track. With crossfade
off, the new track waits in silence and starts the moment the tail ends.
Both are configurable under Settings -> Transitions and persist to
localStorage. Preload is on and crossfade is 400ms by default.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The reprocess_artists case was 208 of index.ts's 515 lines — 40% of the file
and most of what this PR changed in it, buried inside a switch. index.ts is
now 313 lines and reads as what it is: wiring, the job switch, cron
registration, shutdown.
The move also collapses a real duplication. The artist merge and the
normalized_name dedup pass ran the same five statements in the same order
against different id pairs, so the withTransaction change had to be made
twice, identically. Both now call one mergeArtistInto(client, keepId,
loserId), which takes a Queryable so the caller owns the transaction, and the
ON DELETE CASCADE hazard is documented once instead of twice.
Verified as behaviour-preserving: the statement sequence diffs identical
against the previous commit, and workers typecheck is clean.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
db.service.ts was 2085 lines, of which ~700 were not behaviour at all: the
migration registry, the row-shape interfaces, and the column allowlist. That
makes the file painful to review — the reviewer's note on the MIGRATIONS
array.
Three pure moves into backend/src/db/, which already owns schema.sql:
- migrations.ts (513) — the registry, plus a named Migration type
- types.ts (168) — the row shapes
- updatable-columns.ts (42) — UPDATABLE_COLUMNS + allowedFields
db.service.ts drops to 1378 lines and re-exports ../db/types.js, so existing
`import { Track, ListenerBelief } from '../services/db.service.js'` in the
routes, generators and session-director keeps working untouched.
No behaviour change, and verified as such rather than asserted: the migration
id list and the entire 502-line SQL body diff byte-identical against the
previous commit, backend typecheck is clean and 33/33 tests pass.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
REVIEW-2026-07-30.md is the source for the preceding commits. AUDIT.md was
its superseded predecessor.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The dislike lifecycle promised WARNED -> HIDDEN -> deleted, but nothing ever
removed a file: cleanup.service logged its intent behind
MUZICK_ALLOW_HARD_DELETE and returned, and the two backend delete paths
(hardDeleteTrack, permanentlyDeleteTrack) disagreed about what deletion
meant. The review recommended dropping hard deletion and making HIDDEN
terminal; the owner chose to make deletion real instead.
- cleanup.service performs a true unlink() — no trash directory — for
tracks that have been HIDDEN for a 7-day grace period, then settles the
row. This is the single unlink() call site in the system.
- permanentlyDeleteTrack is the one delete path; hardDeleteTrack is gone.
- a deleted_permanent audit row records what was removed, and
migration 20260730_hard_delete_audit_trail backs it.
MUZICK_ALLOW_HARD_DELETE remains OFF: the docker-compose entry is commented
out, there is no enabling default in code, and the worker's /music bind is
the only writable one. Deletion stays dry-run until the owner opts in
deliberately.
REVIEW-2026-07-30.md open decision: dislike lifecycle.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The sweep stats every track path and marks unreadable files missing, with no
check that /music is mounted. An unmounted or misbehaving bind would fail
every stat and mark the entire library missing in one pass; the ratio is only
recoverable by a full rescan.
Three guards, cheapest first:
- liveness: probe a sample of existing track paths before doing anything;
abort if none are readable
- ratio: abort mid-sweep if the missing fraction crosses a threshold,
leaving already-marked rows alone rather than rolling back a partial pass
- progress: keyset pagination over id with the cursor persisted in
integrity_sweep_state, so a sweep aborted or restarted mid-run resumes
instead of re-walking from the top and re-marking
The repair-corrupted-metadata script shares the same failure mode and gets
the same abort path.
REVIEW-2026-07-30.md secondary finding: integrity sweep has no mount check.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The worker ran every job through a single pg Client while BullMQ was
configured with concurrency: 10. A Client is one connection with one
protocol stream and no queueing: ten concurrent jobs interleave on it, and
any BEGIN/COMMIT is shared by all of them, so an unrelated job's failure can
roll back another's work and a rollback can discard a third's committed
intent.
Switched to a Pool, added a small withTransaction(pool, fn) helper that
takes a dedicated connection per transaction, and threaded a Queryable
interface through the services so they accept either a pool or a pooled
client. Both reprocess_artists merge blocks — the artist merge and the
duplicate-album merge — now run inside withTransaction; previously a failure
partway through left artists merged and their tracks unmoved.
integrity.service and cleanup.service get only the constructor type change
here so this commit compiles; their own fixes follow in the next two
commits. cleanup.service's BEGIN/COMMIT-on-a-Pool is therefore still wrong
at this commit and is replaced wholesale by the hard-delete commit.
REVIEW-2026-07-30.md finding 4 (and the concurrency note in finding 3).
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
throttle() read lastRequestAt, awaited delay(), then wrote it back — a
TOCTOU race with no mutex or queue, under concurrency: 10. Ten jobs read the
same timestamp, slept the same duration and fired in the same tick, giving
up to ~10 req/s against MusicBrainz's 1 req/s policy and risking an IP
block.
Replaced the lastRequestAt map with a per-host { lastRequestAt, tail }
limiter; each call links onto that host's promise chain, so the
read-sleep-write critical section is serialized and N concurrent callers
space out by minIntervalMs. Chain rejections are swallowed so one failure
cannot poison the queue. Per-host rather than global, so other integrations
are not starved by MusicBrainz.
Measured: 5 concurrent same-host calls at 200ms -> 802ms (previously all in
one tick); 3 distinct hosts at 1000ms -> 0ms, confirming no cross-host
starvation.
musicbrainz.client caught HttpError and returned null at all 7 catch sites,
making a rate-limited MusicBrainz indistinguishable from "no data for your
library" while every job reported success. A shared logMbFailure() now logs
429 (and 503 whose body mentions a rate limit) at error, stating results are
INCOMPLETE. The error model is otherwise unchanged.
REVIEW-2026-07-30.md finding 8.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
next() did `queue.slice(idx + 1)`, so the current track was always
queue[0]. prev()'s `idx > 0` guard could therefore never pass after an
auto-advance — Previous did nothing, ever — and repeat: 'all' jumped to
queue[0], which is the track that just finished, looping the last track of
an album instead of restarting it.
Replaced with a currentIndex cursor; the queue is no longer trimmed behind
the playhead. The old slice did serve a purpose — bounding Vibe-prefetch
growth — so that is preserved as a MAX_HISTORY = 50 cap that drops the
oldest entries and re-bases the index, rather than dropped outright.
setQueue/playTrack/setCurrentTrack recompute the cursor, next()/prev() fall
back to findIndex if it drifts, and shuffle now picks by index so the
cursor stays valid.
Consumer audit: NowPlayingPanel and Vibe.tsx already derived position via
findIndex and needed no change. TrackRow.handlePlay did
`setQueue(queue.slice(index))`, which re-broke prev at the point of click
even with the store fixed; it now passes the intact queue.
This commit also includes a pre-existing uncommitted fix from the working
tree (not authored by Claude): the end-of-queue auto-resume loop, which
stops playback at the end of the queue instead of restarting. It is correct
and independent of the cursor bug, and is preserved verbatim here.
REVIEW-2026-07-30.md finding 7.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The entire admin surface of the SPA had been dead since auth landed
(5ed8d9e / 3bc9f2d). nginx.conf.template injected only
`Authorization: Bearer ${MUZICK_API_KEY}` for all of /api, docker-compose
passed only MUZICK_API_KEY to the frontend container, and app.ts requires
token === adminKey for /api/admin/*. The two keys differ, and
services/api.ts sets no headers of its own.
All 10 admin call sites were affected: the Jobs page polled 403s every
3s/5s forever and rendered a blank Overview with no error state, and every
Settings library action (Scan, Reindex, Reprocess artists, Re-enrich,
Duplicates merge) silently failed.
Three changes, each necessary:
- a `location /api/admin/` block injecting the admin key
- the Dockerfile envsubst list widened to include MUZICK_ADMIN_KEY,
without which the new variable substitutes to empty and the header
becomes a bare "Bearer"
- MUZICK_ADMIN_KEY passed to the frontend service in docker-compose
nginx selects the longest matching prefix regardless of block order;
verified empirically in a throwaway nginx:stable-alpine running the real
envsubst output against a stub that echoes $http_authorization:
/api/admin/queue-stats -> Bearer ADMINKEY456
/api/admin/duplicates/merge -> Bearer ADMINKEY456
/api/tracks -> Bearer APIKEY123
/api/health -> Bearer APIKEY123
All 10 call sites use /admin/... under the axios /api baseURL and none
request bare /api/admin without a trailing slash.
Also gives the Jobs page an error state: a banner that names a 401/403 as a
missing or wrong admin key, a Retry button, "Loading queue stats..." in
place of a blank Overview, and refetchInterval returning false once the
query has errored so it stops hammering a failing endpoint.
Deletes frontend/nginx.conf — unreferenced by the Dockerfile (confirmed by
grep) and the insecure variant of the template.
Worth noting and not addressed here: the outer LAN-only proxy already
forges credentials for everything reaching /api, so this key split buys no
real security while having cost the whole admin surface. Collapsing to one
key would be simpler.
REVIEW-2026-07-30.md finding 2.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
db.service inserted the 'deleted_permanent' feedback row and then deleted
the track, but feedback.track_id was ON DELETE CASCADE (verified on the
live DB: confdeltype = 'c'), so the audit row deleted itself. feedback
contains zero deleted_permanent rows.
feedback is an audit log and must outlive its subject: the FK becomes
ON DELETE SET NULL. track_id was already nullable, and nothing in backend/
or workers/ SELECTs from feedback — the only other reference is
mergeTracks()'s UPDATE feedback SET track_id, which re-points to the
survivor — so no caller assumed non-null.
Migration 20260730_feedback_track_id_set_null drops the constraint by
matching confdeltype rather than by name, since the live schema has
drifted. Verified on a scratch PG16: confdeltype flips 'c' -> 'n' and a
deleted_permanent row survives its track's deletion.
Correct under either resolution of the dislike-lifecycle decision, so it
lands independently of it.
REVIEW-2026-07-30.md finding 6 (cascade only).
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
updateTrack/updateArtist/updateAlbum built their SET clause from
Object.keys(data) where data is `request.body as any`, interpolating
request-supplied keys straight into SQL as quoted identifiers:
fields.map((f, i) => `"${f}" = $${i + 2}`)
A crafted body key closes the quoted identifier and injects into the SET
list. Mitigated in practice only by the LAN/VPN-only proxy — which supplies
the auth token automatically, so any device on the LAN could reach it from
a browser.
Adds per-table UPDATABLE_COLUMNS plus an allowedFields() helper, applied in
all three methods. The allowlist lives in the service layer rather than the
routes so it covers every caller.
Unknown keys are dropped rather than rejected: the three routes do no error
mapping, so a throw surfaces as a bare 500, and the pre-existing "No fields
to update" error still fires for a payload rejected in its entirety.
Also closes plain mass-assignment. Excluded: path/hash/mtime
(scanner-owned; path is the only link to the read-only bind),
state/quarantined_at/deleted_at (dislike lifecycle and integrity sweep),
play_count/skip_count/dislike_count/last_played_at (learning signal —
forgeable counters poison the engine), and identity/generated columns.
The only callers are the three HTTP PUTs; the frontend's update* service
exports are dead code, so nothing relied on writing an excluded column.
REVIEW-2026-07-30.md finding 5.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
claims declared UNIQUE (..., source, user_id). Every objective claim has
user_id IS NULL, and under default NULLS DISTINCT semantics Postgres treats
those rows as unique, so the ON CONFLICT DO UPDATE / DO NOTHING clauses in
db.service and mb-spine-writer never fired. Re-enrichment inserted a fresh
duplicate every run instead of reinforcing.
Live data: 236 duplicate groups, 2110 excess rows, worst single claim 86
copies, ~15% of 13,910 claims. claim_fusion is SUM(trust * confidence *
recency), so one edge could carry 86x its intended weight — the likely
cause of repetitive recommendations, and almost certainly the root of
d497588 (claim_fusion MV duplicate-key failure).
Migration 20260730_claims_dedup_nulls_not_distinct, two phases in one
transaction. Dedup MUST precede the constraint or adding it fails.
Phase 1 collapses each group into its most recently reinforced row,
carrying forward MAX(last_reinforced_at), MAX(evidence_at) and
MAX(confidence) — reinforcement recency would otherwise be lost by simply
deleting extras. The MAX(...) OVER grp and ROW_NUMBER() OVER ordered
windows are deliberately separate: an ORDER BY inside the window makes the
default frame UNBOUNDED PRECEDING TO CURRENT ROW, which turns MAX() into a
running maximum and would silently keep the wrong confidence.
Phase 2 drops the old constraint by matching its definition rather than its
name, because the live DB has drifted and its autogenerated name is
truncated at 63 characters.
Verified on a scratch PG16 seeded with the old schema plus a 3-row
duplicate group, a distinct-source singleton and a real-user_id row:
UPDATE 3 / DELETE 2, keeper retained the group max of each field from three
different rows, re-run is a no-op, and a subsequent ON CONFLICT DO UPDATE
with user_id = NULL fires correctly.
REVIEW-2026-07-30.md finding 4.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The CTE was `WITH halflives AS (SELECT profile, CASE profile ...)` with no
FROM clause. Postgres rejects it with 42703 (column "profile" does not
exist) on every hourly invocation, so the temporal dimension of the
recommendation engine had never executed once — obsession (14d half-life)
and contextual (7d) never faded.
Rewritten as `WITH halflives(profile, halflife_sec) AS (VALUES ...)`,
half-lives preserved exactly.
One deliberate semantic change: the broken CASE had an `ELSE 30 * 86400`
fallback, so an unrecognised profile would have decayed on a 30-day
half-life. The VALUES join leaves unknown profiles undecayed instead.
Today that is a no-op (only `forgotten`, already excluded by the WHERE),
but a future profile added without a half-life will now conspicuously not
decay rather than quietly decaying at an arbitrary rate.
Verified against a scratch PG16 with one belief per profile aged exactly
one half-life: UPDATE 2, obsession and contextual halved, forgotten and a
fresh longterm untouched. Against the live DB (in a rolled-back
transaction) the fix reports UPDATE 843.
NOTE ON ROLLOUT: the first successful run applies ~23 days of accrued
decay at once, cutting obsession beliefs to ~0.32x. That is correct
behaviour, but recommendations will shift visibly. Expected, not a
regression.
REVIEW-2026-07-30.md finding 3.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Three separate insert paths made a fresh Postgres volume unusable. The live
database only works because its volume predates the constraints.
- scanner.service.resolveOrCreateArtist inserted only (name), but
schema.sql declares canonical_name NOT NULL with no default. Every
artist insert failed, and processFile swallows per-file errors, so a
scan reported success with 0 tracks and a permanently empty library.
- enrichment.service inserted explicitly into artists.normalized_name,
which is GENERATED ALWAYS AS (normalize_artist(name)) STORED:
"cannot insert a non-DEFAULT value into column" (428C9). All
enrichment artist creation failed on a fresh volume.
- db.service.createArtist omitted canonical_name, same failure.
canonical_name holds the raw tag name, not normalize_artist() output,
which truncates on `/` and a standalone `x` ("AC/DC" -> "AC"). That is the
convention createLocalArtist already used. The truncation bug in
artists.name is pre-existing and deliberately left untouched here.
Verified on a scratch postgres:16-alpine with the real schema: the old
statement reproduces the NOT NULL violation, the new path yields
artists/albums/tracks/track_artists rows.
REVIEW-2026-07-30.md finding 1.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The MV SELECTed COALESCE(user_id, zero-uuid) but GROUPed BY raw user_id,
so a global enrichment claim (user_id NULL) and a default-user behavior
claim (user_id = zero-uuid) for the same edge produced two rows that
collide on idx_claim_fusion_pk, breaking REFRESH ... CONCURRENTLY.
Exposed by the #219 listener_behavior same_scene_as/alias_of writes.
New migration rebuilds the MV grouping by the COALESCE'd user_id so the
two fuse into one row.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Per-user SET NX PX lock with Lua CAS release around the
getActivePlan->mutate->setActivePlan span so concurrent prefetch +
feedback requests serialize instead of losing one side's write.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
kami
kami