Helm session recovery: unshipped fixes, unconfirmed root causes, and cleanup items (2026-08-19)

[tanflem]

Summary

Recovered from ~28 old helm/* session transcripts (worktrees auto-deleted, session logs survived under ~/.claude/projects/). A peer session (help-me-get-a-tight-list-of-what-i-need-to-do-to) had already done most of this recovery ~23h earlier and opened PRs #9502–#9506 from it. This issue captures what's left: things still unticketed, a likely-duplicate PR, and one open architecture decision.

Needs a decision

• PR feat(api-media): activate video variant reconciliation (QA-554) #9386 architecture question — should availableLanguages on a parent video stay a mutated cache, become a plain view, a materialized view, or a query-time cache? A prior session had already ruled out "plain view" and "query-time cache" as non-viable for arclight's paginated/filtered catalog queries when it went idle mid-discussion. Blocks feat(api-media): activate video variant reconciliation (QA-554) #9386 and feat(videos-admin): add processing view (QA-554) #9384, which block VMT-318.
• PR feat(api-media): durable, retriable video-to-Algolia sync (QA-543) #9504 (durable, retriable video-to-Algolia sync (QA-543)) — looks like a duplicate of already-merged PR fix(api-media): make video Algolia indexing durable and retriable (QA-543) #9442 (QA-543 shipped 2026-08-10). Opened 2026-08-18 from the stale helm/qa-543-durable-video-algolia-sync branch, probably without realizing the work already landed. Worth closing after confirming the diff is genuinely redundant.
• PR feat(videos-admin): add processing view (QA-554) #9384 — has an active merge conflict. A prior triage session flagged that GitHub's diff view for this PR was seen showing a stale/flipped-base 10-file diff vs. the true 38-file/+2343-line diff (includes a Prisma migration) — worth double-checking the diff view is accurate before anyone reviews it.

Confirmed root causes, not yet fixed

• Ama / bcp47-less-language 404 on jesusfilm.org/watch (whole bug class, not Ama-specific) — root cause confirmed in the forge repo, not core. The watch-page language picker (apps/web/src/lib/content.ts in forge) lists any published+hls language, but the route gate rejects any slug missing from a generated corpus (packages/watch-url-policy/src/public-watch-language-slugs.ts in forge) built from Language.bcp47. ~39 admin-added Language rows (including Ama, id 20118, bcp47 nyi) have no bcp47 set and are skipped by the corpus generator, so classify() returns unknown and notFound() fires before any GraphQL lookup runs. Every published dub in one of those ~39 languages 404s identically. Three fix options identified: backfill bcp47 for the 39 rows, key the corpus generator on Language.slug instead, or filter the picker to match the corpus. Write-up is in core PR docs(solutions): add investigation notes for the Ama watch-page bug #9506 (docs only) — the actual fix needs to happen in forge.
• hasVideos never set true on the Mux-upload path — createOrUpdateVideoVariant (apis/api-media/src/schema/videoVariantUpload/service.ts) never calls ensureLanguageHasVideosTrue, unlike the GraphQL-mutation path. Fix already written and committed locally on an unpushed branch (08-05-TF-fix-language-has-videos-upload-path, commit 950eee08e), needs a rebase onto current main; overlaps with PR feat(api-media): activate video variant reconciliation (QA-554) #9386's reconciliation work.
• Darkroom Faith replacement episodes not playing on Watch — root cause found via a live diagnosis session (14/24 variants affected, deterministic repro): production Watch reads from a separate coreId-keyed content mirror that isn't in this monorepo — same shape as the Ama bug above. A full ranked-hypothesis handoff doc exists only at /tmp/handoff-darkroom-faith-watch-20260812/HANDOFF.md in an ephemeral container and will be lost if not rescued.
• Possible cache cross-contamination between clients — apis/api-media/src/yoga.ts keys anonymous-request cache entries on null session instead of clientName, so an anonymous journeys response could theoretically be served to a watch request. Means the shipped restrictViewPlatforms feature could already be leaking in prod via cache cross-contamination. (Already privately tracked as issue add first-draft journey graph design #6 in a personal core-issues repo — flagging here too since it's a real prod-facing risk.)
• Prod Redis/Arclight languageIds bug — some caller passes a BCP-47 tag (en) into Arclight's languageIds param (which expects numeric IDs), producing and caching an empty result for English media-components. One stale cache key was deleted live as a stopgap (media-components:1:4999:languageIds:en). The actual caller sending the malformed request was never identified.
• Arclight download URLs hardcoded to a ≤720p tier — apps/arclight/src/lib/downloadHelpers.ts's findDownloadWithFallback only ever looks up VideoVariantDownloadQuality.low/.high rows, never fhd/qhd/uhd/highest. Every video's download-URL API field is capped regardless of actual encode quality. HLS streaming through Arclight is unaffected — only the download-URL field.
• api-languages response cache has no TTL and no invalidation path for importer-written data — could cause indefinitely stale language lists after an import.
• LanguageAutocomplete.tsx:71 silently drops languages whose name array is empty, with no warning surfaced anywhere.
• forge issue chore: upgrade to node 18 #1831 ("public resolvers don't filter DRAFT/unpublished status") — still open, surfaced while diagnosing and fixing the 2_ElCamImpulsesVert restricted-video leak (that fix itself shipped, forge PRs fix: make journeys published by default #1829/fix: journeys variant language #1830/feat: ui elements improvements #1838/fix: permission error #1839, all merged).

GitHub issue triage conclusions still current (from an old /triage pass)

• Dedupe the two Algolia client factories in api-media #9434 (dedupe Algolia getRequiredEnv factories) — recommended ready-for-agent, still open, still unlabeled. Self-contained pickup.
• Allow unpublished videos to be used in journeys-admin #9464 (unpublished videos in journeys-admin) — was blocked on fix(journeys-admin): resolve unpublished variants in the video block picker #9430/fix(arclight): filter unpublished variants from media-component languages endpoints #9429; both are now merged, so this just needs closing/verifying.
• Unpublished-video warning UI in journeys-admin (banner + confirmation modal) #9465 (unpublished-video warning UI) — still open, blocked on copy from Jessie.
• Communicate unpublished-video exposure to videos-admin users #9466 (comms to videos-admin users) — still open, recommended splitting into comms/UI halves, never done.

At risk of being lost (git-gc pruning)

• A 293-line regression runbook (docs/qa/regression-runbook-2026-08-11.md, covering 28 merged PRs) and an already-written VideoList.tsx row-click bugfix (URL built from pathname instead of a fixed path, broken by PR feat(videos-admin): add status pipeline tab #9361 moving the component) both live only in a dangling, unreferenced commit (a24436b0e in the main repo's object store) — not on any branch, so eligible for git gc pruning. The runbook is also still live as a Claude Artifact (https://claude.ai/code/artifact/f9f55fd3-d1aa-4a5c-93ab-8e03908d7b69) as a backup, but the bugfix commit has no other copy.
• A prod resync script (resync-restrict-view-watch.ts, already run successfully against prod — 46/46 videos updated, this part is done) is recoverable from a stash (stash@{0} in the main repo, "wip before checking out PR 9430 for testing", commit 9dffcf0a8) if the tooling itself is worth keeping around for reuse.

[tanflem]

This was generated by AI during triage.

Triage update — three of the items in this issue are now resolved

PR #9504 — closed. Confirmed the duplicate you suspected. 19 of its 22 files are byte-identical to main (the whole workers/videoAlgoliaSync/ directory included); QA-543 landed via #9442 on 2026-08-10. The three files that differ do so because main moved further ahead — both of the branch's real changes (enqueueVideoAlgoliaSync at the updateAvailableLanguages call site, and the findContainerParentIds extraction) are already on main. Merging it would have regressed main, which has since replaced addLanguageToVideo's read-then-set with an atomic $executeRaw UPDATE to close a lost-update race; the branch still carries the racy version.

PR #9384 — merge conflict resolved. The diff-view discrepancy noted here is explained: the branch carried its own squash-merged foundation (#9385), so main and the branch had add/add conflicts across all of videoVariantReconciliation/ plus audit-parent-variants.ts. Rebuilt on top of #9386's branch and retargeted to it, so the stack is explicit. Now 7 files / +333 / −5, down from 192 files / +9334. 7/7 tests pass.

PR #9505 — closed as already fixed on main. All three commits redundant: hasValidRenditionMetadata is on main (processing.ts line 68), git dropped the lint commit as "already upstream", and the anti-loop guard targets an un-paginated while loop that main replaced with a cursor-paginated one — the bug no longer exists.

Also done

PR #9503 — rebuilt on top of #9507 and retargeted to its branch. These two were divergent implementations of the same parentVariantsOnly work with different patch-ids, so whichever merged first would have broken the other. #9507 is now the single source; #9503 carries only the #9468 audit tooling. 40/40 then 439/439 tests pass. #9507 must merge first.

PR #9502 — recut onto current main (was 208 commits behind, cut from 2026-05-18). 14/14 tests pass.

Still open, deliberately

The #9386 architecture question — whether availableLanguages on a parent stays a mutated cache, a materialized view, or something else — is untouched and still blocks #9386 → #9384 → VMT-318. That is the one item here that needs a human decision, and it is now the critical path for the whole QA-554 stack.

No issues were closed. #9468, #9382, #9483 and #9434 all still have unmerged or unstarted work — #9382's zero-size-download half has shipped but its parent-language-Variant half is #9507, still in review.

[tanflem]

This was generated by AI during triage.

Triage update — 2026-08-21

One of the three "Needs a decision" items has resolved itself; the other two stand.

Resolved — PR #9504 (duplicate of merged #9442): confirmed redundant and closed on 2026-08-19. No action left.

Still open — PR #9386 architecture question: unchanged. The availableLanguages mutated-cache vs. view vs. materialized-view decision is still unmade, and still blocks #9386 and #9384 (and through them VMT-318). Note that #9515 and its sub-issue tree (#9517–#9525) have since been opened and are actively shipping against the mutated-cache option — nine PRs in flight (#9528–#9535). If that is the de facto answer, this item can be closed as decided-by-implementation; if not, the in-flight work is building on a base that is about to be reconsidered. Worth an explicit call either way.

Still open — PR #9384 merge conflict: unchanged, still conflicted. It also targets tannerfleming/qa-554-reconciliation-activation (the head of #9386), not main — so the stale/flipped-base diff view flagged in the issue body is at least partly explained by the stack, and the diff should be read against #9386 rather than main.

Also since this issue was written: #9464 closed (shipped via merged #9430), #9492 closed (obsolete — Phoebe's branch scheme now passes the existing pattern), and PR #9532 closed as a duplicate of #9534.