fix(playback): warm catalog prefetch on detail reads, retry slow transcode manifests

[drondeseries]

Problem

Related issue: N/A
Validation tasks: none

Cold virtual starts still pay a full 3-7s provider list: clients press play straight from catalog and season lists, and the watch-detail prefetch from the earlier change never fires because no client reads watch detail first. Separately, a slow-but-alive transcode encoder answers a terminal-looking 503 while FFmpeg is still producing, and the web player burns its fatal budget instead of polling again.

Approach

Warm the same bounded best-result cache from the reads clients actually make: item detail warms its versions, season-episode lists warm the first episode's files only. A file-level admit gate plus a 4-slot load semaphore keep high-volume catalog reads from fanning out row loads before the existing bounded prefetch queue.

Manifest polls distinguish slow-alive from dead: alive-but-slow answers retryable 503 (not_ready_retry + Retry-After), and the web player reloads the manifest source while startup is still in flight instead of counting it against fatal recovery. Post-startup failures stay on the fatal path.

Validation

• go build ./..., gofmt, go vet: clean
• Focused handler tests (prefetch identity + collapse, decode-rejected, manifest): pass, including -race on prefetch
• Web: pnpm run lint 0 errors, prettier clean, new manifest-not-ready + guard suites pass
• Oracle adversarial review across 5 rounds to APPROVED (poll-back restricted to startup state, semaphore race fixed with sync.Once, invalid-200-stub rejected in favor of retryable 503)
• Full CI required on the head
• Android manifest retry behavior unverified follow-up

Risks

Prefetch stays speculative: first-2 file IDs, virtual-only filter, shed-when-saturated. Catalog reads gain a bounded background load, never response latency. Manifest not_ready_retry is a new code web clients poll on; other clients see a plain 503 with Retry-After as before.

Before / after

Signal	Before	After
Cold start from catalog	Full 3-7s list, nothing warmed	Detail/season reads warm up to 2 rows ahead of play
Repeated catalog reads	Each pays row loads	Same rows collapse pre-load; distinct keys bounded to 4 concurrent loads
Slow transcode manifest	Bare 503 treated as terminal	Retryable 503; web polls again while starting, fatal path after

Checklist

• I read and can explain the complete diff.
• This pull request addresses one concern.

AI Disclosure

• Harness: OpenCode
• Tool(s): OpenCode default tools; explorer and oracle subagents for scouting and review
• Model(s): omniroute/workhorse (Workhorse, Gemini 3.8+ Tiered with Failover)
• Involvement: AI-assisted
• Adversarial review: oracle reviewed across 5 rounds (invalid playlist stub, admission-before-load, client retry behavior, semaphore init race, startup-state gating). All findings fixed and re-reviewed to APPROVED.