commit b0c16f2af9326934652dd4591cff6d3b4391f0ed
parent b2ad3270da6d3b219fc7466b059617dbf73e4576
Author: I Mean I'm Just Saying <imeanimjustsaying@kiwifarms.st>
Date: Wed, 29 Jul 2026 11:43:19 -0400
Correct the roadmap's gate list, and record what the measurements changed
Three corrections the exploration and the measurements forced:
- **Phase 1.5 is not a backfill gate.** `digestContext-server.ts:31-35` hashes
the EMPTY note to a stable value, so adding a channel note later invalidates
only that channel. What would invalidate the corpus is bumping the
`digest-context-v1` prefix — so the gate is a DECISION to take before
sweeping, not work to do first. PLAN.md said the opposite in two places.
- **"Do not backfill before 2.5 and 11a" named the wrong things.** Both were
largely landed. The things actually missing were never on the list: no
launcher, no boot resume, no pause writer, no GPU arbitration.
- **`updateDuplicateOverride` had zero callers**, so a fails-closed gate had no
way to be opened and ~166 clusters could never be confirmed by anyone. Worth
generalising, and recorded as such: "the extension point exists" is not the
same as "the feature works".
FACTS.md gains the four-way threshold bracket with its marginal-band tables, the
audio-hour pricing of the sweep at each threshold, the heaviest-channel work
queue, and the 14.5% id-vs-directory measurement.
The correction that matters most is to a claim these docs made themselves:
duplicate sharing is worth ~1.7 sweep days of 80, not "~11% of the sweep",
because cluster members are 19% of the corpus by video count and 4.3% of its
audio-hours. The stale "consequence for the digest backfill" paragraph is marked
overstated rather than deleted — it was right about pairs and wrong about time,
and that distinction is the reusable lesson.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Diffstat:
| M | PLAN.md | | | 32 | +++++++++++++++++++++++++------- |
| M | plans/FACTS.md | | | 97 | +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++---- |
| M | plans/STATE.md | | | 142 | +++++++++++++++++++++++++++++++++++++++++++++++++++++++++---------------------- |
3 files changed, 221 insertions(+), 50 deletions(-)
diff --git a/PLAN.md b/PLAN.md
@@ -246,7 +246,10 @@ Markdown with optional frontmatter: `hosts`, `recurring_guests`,
inference silently degrade every downstream summary for a channel, invisibly, because the
output would still look plausible.
- Hash the resolved context into `ai-digest.json.contextHash` so editing notes correctly
- marks digests stale. **This is why backfill must not start before 1.5 exists.**
+ marks digests stale. **Corrected 2026-07-29:** this is NOT why backfill must wait. The
+ empty note already hashes to a stable value, so a note added later invalidates only its own
+ channel. The corpus-wide risk is the deliberate one — bumping the `digest-context-v1` hash
+ prefix — so what must precede a sweep is that DECISION, not this phase.
### Phase 2 — Digests as a first-class corpus — **DONE (2026-07-29)**
@@ -573,16 +576,31 @@ browser only to diagnose failures.
### Ordering traps
-- **Do not backfill before 1.5.** Channel context feeds `contextHash`, which correctly
- invalidates every digest generated without it. At 15–60 s per video that is a costly redo.
-- **Do not backfill before 2.5 and 11a.** An unobservable multi-day sweep cannot be tuned or
- safely interrupted, and one with no review path is a one-shot gamble on prompt quality.
+- **CORRECTED (2026-07-29): 1.5 is not a backfill gate — a hash-prefix decision is.**
+ `digestContext-server.ts:31-35` hashes the EMPTY note to a stable value, so adding a
+ channel note later invalidates only that channel, not the corpus. What would invalidate
+ everything is bumping the `digest-context-v1` prefix (`:42`), which 1.5-as-specced does.
+ So the gate is a decision to take before sweeping, not work to do first.
+- **CORRECTED (2026-07-29): "before 2.5 and 11a" named the wrong things.** Both were largely
+ landed — the metric union, `METRIC_PREFIX`, per-job progress and the `noDigest` bucket all
+ existed. The things actually missing were never on this list: **no corpus-wide launcher,
+ no boot-time resume, no pause writer, and no GPU arbitration**. All four are now built
+ (`controller/digestSweep.ts`, `controller/digestYield.ts`, `instrumentation.ts`,
+ `DigestSweepControls`). The remaining honest gate is the review path, and its minimum —
+ persisted failure warnings plus a `digestWarnings` bucket — is in.
- **Benchmark before committing to backfill.** The smoke test measured 15.5 s for a ~2.7k
token chunk; a 176-minute podcast needs roughly a dozen chunks. Budget minutes per long
video and multiply by corpus size. The remote lane changes this arithmetic — measure both,
then split the backlog between them.
-- **Fix the duplicated metric union before extending `JobProgressMetric`**, or Phase 2.5's
- compiler-driven audit silently misses a consumer.
+- **RESOLVED: the duplicated metric union is fixed.** `RunningJobsList` imports the union
+ and `MonitorWidget` uses a `METRIC_PREFIX` lookup table, so extending it is now
+ compiler-checked.
+- **Price a cost lever in AUDIO-HOURS before believing it** (`common/bin/digest-plan.ts`).
+ Measured 2026-07-29: duplicate-cluster sharing is worth **~1.7 sweep days of 80**, not the
+ "~11%" `digestSharing.ts`'s header claims — cluster members are 19% of the corpus by video
+ count but **4.3% of its audio-hours**, because mirrors skew SHORT and the sweep is
+ dominated by unclustered long-form VODs. GPU contention with whisper is worth ~56 days on
+ the same measurement, roughly fifty times more than every duplicate lever combined.
### Hardware and prerequisites
diff --git a/plans/FACTS.md b/plans/FACTS.md
@@ -490,15 +490,104 @@ timing-aligned** enough for a shared digest to be placed correctly — **2.5× t
1,599 previously banked**. Against a corpus of 76,318 videos that is ~5.2% of the
sweep avoidable by sharing rather than ~2.1%.
-**Not yet changed:** `NEAR_THRESHOLD_DEFAULT` still ships at 0.6. Lowering it
-changes what every built site asserts to readers, so it is recorded here as a
-measured recommendation rather than applied as a side effect of a digest task.
-See STATE.md.
+**CHANGED 2026-07-29: `DEFAULT_NEAR_THRESHOLD` now ships at 0.35**, in its own
+commit, after bracketing it at four values — see the next section. Note that the
+"consequence for the digest backfill" above was **overstated**: the saving is
+real in pair count but nearly absent in audio-hours, which is the unit the sweep
+is actually priced in. ~5.2% of pairs is ~1.7 sweep days of 80.
Reproduce with:
`pnpm exec tsx bin/duplicate-shorts.ts --all-durations --blocking title --near 0.35`
(back up `transcripts/duplicates.json` first — the run overwrites it).
+### Bracketing the near threshold — 4 runs, identical inputs (measured 2026-07-29)
+
+Corpus-wide, `--all-durations --blocking title`, only `--near` varied. 76,354 videos
+scanned, 8,352 pairs nominated in every run.
+
+| `--near` | confirmed | rejected | clusters | videos | aligned | needsReview | wall |
+| --- | --- | --- | --- | --- | --- | --- | --- |
+| 0.6 | 2,736 | 5,404 | 2,846 | 5,746 | 4,273 | 168 | 124 s |
+| 0.45 | 7,147 | 993 | 7,110 | 14,344 | 10,747 | 166 | 199 s |
+| **0.35** | **7,483** | **657** | **7,434** | **14,997** | **11,217** | 166 | 225 s |
+| 0.25 | 7,542 | 598 | 7,493 | 15,115 | 11,297 | 166 | ~230 s |
+
+Every step down is a **strict superset** — 0 videos lost at any step. Returns collapse:
+0.6→0.45 admits **4,264** entirely-new clusters, 0.45→0.35 admits **324**, 0.35→0.25 admits
+**59**.
+
+The marginal bands were read, not counted:
+
+| band | new clusters | byte-identical title | runtime ±2 s | cross-platform | SAME-channel |
+| --- | --- | --- | --- | --- | --- |
+| 0.6 → 0.45 | 4,264 | 96.2% | 91.1% | 99.5% | **3** |
+| 0.45 → 0.35 | 324 | 97.2% | 83.3% | 96.0% | **2** |
+| 0.35 → 0.25 | 59 | 94.9% | 86.4% | 93.2% | **0** |
+
+The riskiest cases in each band were printed individually. In the 0.6→0.45 band they are all
+YouTube↔Rumble pairs agreeing on title (modulo whitespace), runtime to the second, and upload
+date. The 0.45→0.35 band contains the only two arguable cases in the whole sweep: a
+`HasanAbiVODs` pair with the same title but runtimes 9 minutes apart, and an `omnibased` pair
+with identical title/duration/date. Neither is a clear false positive; the first would be
+refused by the alignment gate regardless.
+
+**Adopted: 0.35.** The recall knee is at **0.45** — the value to take if a more conservative
+assertion is ever wanted.
+
+### The digest sweep, priced in AUDIO-HOURS (measured 2026-07-29)
+
+`common/bin/digest-plan.ts --no-freshness`, at the measured 90 s/audio-hour. Total digestable
+corpus: **77,298 audio-hours** over 73,367 videos (2,987 have no transcript), which reproduces
+the roadmap's ~81-day headline independently.
+
+| `--near` | canonical | mirror-aligned (free) | mirror-unaligned | unclustered | **to generate** | **days** |
+| --- | --- | --- | --- | --- | --- | --- |
+| 0.6 | 853 h | 494 h | 344 h | 75,607 h | 76,804 h | 80.0 |
+| 0.45 | 3,472 h | 1,448 h | 1,866 h | 70,513 h | 75,851 h | 79.0 |
+| 0.35 | 4,070 h | 1,661 h | 2,206 h | 69,361 h | 75,638 h | 78.8 |
+| 0.25 | 4,137 h | 1,683 h | 2,242 h | 69,236 h | 75,616 h | 78.8 |
+
+**Duplicate sharing is not a meaningful cost lever, and `digestSharing.ts`'s "~11% of the
+sweep" header is wrong.** Cluster members are 19% of the corpus by video count but **4.3% of
+its audio-hours**: mirrors skew SHORT while the sweep is dominated by unclustered long-form
+VODs. Loosening the threshold as far as it goes moves the sweep from 80.0 to 78.8 days.
+
+Note the `mirror-unaligned` column: only ~55% of mirrors pass the alignment gate, so counting
+all mirrors as free would overstate the saving by nearly half. The plan splits them.
+
+Heaviest channels (the sweep's work-queue order), audio-hours to generate at 0.35:
+
+| channel | h | days |
+| --- | --- | --- |
+| HasanAbiVODs3 | 8,331 | 8.7 |
+| omnibased | 7,783 | 8.1 |
+| HasanAbiVODs | 5,380 | 5.6 |
+| rekietalaw | 4,993 | 5.2 |
+| destiny | 4,020 | 4.2 |
+| shondo-vods | 3,332 | 3.5 |
+
+`HasanAbiVODs3` alone outweighs every duplicate mirror in the corpus combined.
+
+### The id-vs-directory mismatch (measured 2026-07-29)
+
+A slug is `${channelSlug}/${id}` and an id is **not** the on-disk directory name for
+**11,175 of 77,106 indexed videos (14.5%)**, concentrated almost entirely in Rumble
+re-uploads:
+
+| channel | dir !== id |
+| --- | --- |
+| the-quartering-rumble | 7,870 (of 7,870 — every video) |
+| leaflit-rumble | 838 |
+| midwestly | 758 |
+| rekietalaw-rumble | 641 |
+| omnimirror | 522 |
+| cornbreadman | 337 |
+
+Verified directly: for all 7,870 `the-quartering-rumble` entries, `stat.slug ===
+channel/id` and `stat.slug !== channel/dir`. `digestBatch` looked the cluster plan up by
+DIRECTORY, so every one of those lookups missed silently and the mirror regenerated. See the
+`videoDir` field on `DuplicateVideoRef`.
+
### What this unblocks
Corpus-wide detection is no longer off. `buildDigestClusterPlan` reads
diff --git a/plans/STATE.md b/plans/STATE.md
@@ -3,27 +3,41 @@
The working memory for the local-AI derived-corpus work. Rewritten at the end of every
session, before context is cleared. See [`README.md`](README.md) for the protocol.
-**Last updated:** 2026-07-29 — **Phases 2 and 3: the digest layer SHIPS.** The 102 digests
+**Last updated:** 2026-07-29 — **the backfill is at the starting line.** A corpus-wide sweep
+can now be started from one control, survives a server restart, can be paused without being
+lost, yields the GPU to transcription, and reports coverage and an ETA in the unit the work
+is actually priced in. Stages A–D of the "cheapest-first" plan are all in; see the decisions
+below for what the measurements changed.
+
+**The headline is a negative result, and it should be read before planning any more cost
+work.** The plan's own premise — shrink the work-list via duplicate sharing — does not pay.
+Priced in audio-hours by the new `common/bin/digest-plan.ts`:
+
+| `--near` | clusters | to generate | sweep days | saved by sharing |
+| --- | --- | --- | --- | --- |
+| 0.6 (was) | 2,846 | 76,804 h | 80.0 | 0.5 d |
+| 0.45 | 7,110 | 75,851 h | 79.0 | 1.5 d |
+| **0.35 (now)** | **7,434** | **75,638 h** | **78.8** | **1.7 d** |
+| 0.25 | 7,493 | 75,616 h | 78.8 | 1.8 d |
+
+Loosening the threshold as far as it goes buys **1.2 days of 80**. The plan assumed mirrors
+skew long; **they skew short**. Cluster members are 19% of the corpus by video count and
+**4.3% of its audio-hours** — the sweep is dominated by unclustered long-form VODs, and
+`HasanAbiVODs3` alone (8,331 audio-hours) outweighs every mirror in the corpus combined.
+Only ~55% of mirrors pass the alignment gate, so half the nominal saving is refused anyway.
+
+**The lever that does pay is GPU arbitration**: 90 s/audio-hour measured against 27 s
+projected on an idle box is ~80 days versus ~24 — about fifty times every duplicate lever
+put together. It is now implemented (`controller/digestYield.ts`) and **still unmeasured in
+production**; measuring it is the single most valuable next action.
+
+**Previous entry —** 2026-07-29 — **Phases 2 and 3: the digest layer SHIPS.** The 102 digests
already on disk now travel generation → LMDB → a shared `/digests/<slug>/` page tree →
compose → the viewer, where a reader can open a Digest panel, click a chapter and seek to it.
Proven end to end against the real corpus, not fixtures: `community-notes/v2cywen` renders its
20 real chapters (first at 01:09:48 — the validation run's worst coverage gap) and the
playhead follows a click. The control is hidden on the 99.9% of videos with no digest, and a
-borrowed digest is labelled as borrowed. Deliberately independent of the three backfill gates
-(1.5 context, 2.5 observability, 11a review), none of which moved.
-
-Verification that matters: common 356 tests (346 + 10 new), `tsc` clean in all three
-packages, export playwright **150/150** (141 + 9 new), a no-op rebuild skips digest pages,
-and a hand-written `ai-digest.overrides.json` produces exactly `~1 changed` and ships the
-corrected title — the `digestMs` max-of-two-sidecars requirement, which nothing else tests.
-
-**Previous entry —** 2026-07-27 — Stage B2: the digest layer is now INSPECTABLE and its defaults
-are the ones the bake-off measured best. Settings gained a real Digest section, each video
-page gained a review panel (provenance, warnings, `derivedFrom`, staleness), the two digest
-counters were made to agree, `chunk-local`/8192/600 shipped as defaults behind a
-`PROMPT_VERSION` bump, the pilot's "un-created job" was root-caused (a pre-hydration lost
-click) and fixed, and a **102-video validation run** was completed on the real corpus. Also:
-the 0.6 duplicate near-threshold was measured and found to be rejecting real mirrors in bulk.
+borrowed digest is labelled as borrowed.
---
@@ -33,9 +47,9 @@ the 0.6 duplicate near-threshold was measured and found to be rejecting real mir
| --- | --- | --- |
| 0 · Benchmark transcription engines | not started | Still unmeasured. The DIGEST bake-off is done; this is the separate transcription one. |
| 1 · Generation harness | **done + validated** | Spine in `8c041fd`; correctness + bake-off + e2e in Stage B1; operator surfaces + measured-best defaults + 102-video validation run in Stage B2. |
-| 1.5 · Channel context | not started | `contextHash` is plumbed and empty, so notes can be added without invalidating the corpus. |
+| 1.5 · Channel context | not started | **Not a backfill gate** — the empty note hashes stably, so a note added later invalidates only its own channel. The gate is the `digest-context-v1` prefix decision. |
| 2 · Digest corpus in build | **done** | Shared `/digests/<slug>/` page tree, `digests` + `digestPageHashes` + `channelDigestStats` sub-DBs at `SCHEMA_VERSION` 13, `digestMs` in the mtime record and the channel signature, compose reconcile, `CORPUS_SPEC_VERSION` 3. |
-| 2.5 · Observability | **partly landed** | Its stated first step is already done — see the correction below. |
+| 2.5 · Observability | **done** | Corpus coverage on the dashboard + widget, `noDigest` counters, a digest instrument, audio-hour ETAs, and the two progress bugs fixed. |
| 3 · Viewer `?vm=digest` | **done** | Not `?vm=summary` — see the naming decision below. |
| 4 · Search indexing | not started | |
| 5 · Auto-queue | not started | |
@@ -44,7 +58,7 @@ the 0.6 duplicate near-threshold was measured and found to be rejecting real mir
| 8 · Visibility policy | not started | |
| 9 · Attribution + quote filtering | not started | |
| 10 · Lead with the derived corpus | not started | |
-| 11a · Review queue | not started | **Land before backfilling.** `warnings[]` is the data it reads. |
+| 11a · Review queue | **minimum landed** | Total failures now persist warnings (they used to persist none), `buckets.digestWarnings` + a `digest_warnings` filter + an /actionable section, and duplicate-cluster confirm/reject. Approve/dismiss state deferred — needs new persistence. |
| 11b · Viewer feedback | not started | |
**Three corrections to this document's own claims, verified against the code 2026-07-29.**
@@ -63,26 +77,26 @@ Exploration for Phase 2 found the planning docs describing more unbuilt work tha
line and every cross-channel surface gets a counter it is already paying for. Left
unchanged here deliberately: it is Phase 2.5's to take, not a digest-corpus side effect.
-**Recommended next**, in the order the validation run argues for:
-
-1. **Phase 2.5 observability, and treat the throughput finding as its first requirement.**
- The run measured **90 s per audio-hour**, projecting **~81 sweep days** against round 2's
- 24.2 — because short videos cost ~2× per audio-hour and the box was sharing a GPU with
- `auto-transcribe`. A sweep of that length needs a coverage instrument and a scheduling
- story (run the digest lane *against* the transcription lanes, not beside them) before it
- starts, not after. Numbers in FACTS.md.
-2. **Phase 11a review queue.** `warnings[]` is now proven as its data source: 93% of all
- rejections are `out-of-range` from a single guard, and the per-video panel already showed
- that the worst coverage gap in the run is a video where 13 proposed chapters were clamped
- away rather than never proposed. That distinction is what a queue must surface.
-3. **Decide the duplicate near-threshold in its own commit.** 0.6 measurably rejects real
- cross-platform mirrors; 0.35 recovers 4,456 clusters with no observed false positives and
- lifts the shareable-digest ceiling 2.5×. Probe 0.25/0.45 to bracket the knee first.
-4. **Phase 1.5 channel context** — still a backfill gate (`contextHash` staleness), and still
- not required for a validation run.
-
-Deferred from Stage B2 and still open: the `/actionable` duplicate-cluster confirm/reject
-buttons (PLAN.md Phase 11a, extension points recorded there) and the PipelineBand instrument.
+**Recommended next**, in the order the measurements argue for:
+
+1. **MEASURE THE GPU YIELD IN PRODUCTION.** Everything else is second by a factor of fifty.
+ Start a transcription job during a digest sweep, confirm the digest lane idles rather than
+ contending, and measure seconds-per-audio-hour with `yieldToTranscription` on and off. If
+ the 2.8x is real the sweep is ~24 days, not ~80, and every other estimate in these docs
+ changes. If it is NOT real, the contention hypothesis is wrong and the 90 s/audio-hour
+ needs a different explanation — which is just as valuable to know before spending 80 days.
+2. **Run the sweep for a day and watch it.** The launcher, resume, pause and coverage
+ readouts are all built and unit-tested but have not driven a real multi-channel run. Kill
+ the editor mid-run and confirm it resumes with zero rework (eligibility is re-derived from
+ disk, so the correct result is zero).
+3. **Decide the `digest-context-v1` hash-prefix question before sweeping.** This is the real
+ 1.5 gate, and it is a decision, not work: bumping the prefix invalidates the entire
+ corpus, so it must happen before 80 GPU-days go in, or not at all.
+4. **Then Phase 4 search indexing / Phase 7 tags,** which are starved until coverage exists.
+
+Deferred deliberately: Phase 11a approve/dismiss state (needs a new sidecar field or sibling
+file), Phase 1.5 as fully specced, Phase 6 Ollama `/ask` (genuinely independent — a good
+parallel task), Phase 0 transcription benchmark (a separate bottleneck).
---
@@ -93,6 +107,56 @@ relitigate. Earlier entries (fabric deferred, ollama-direct as the structured de
Claude Code on a second queue key, digests in their own page tree, per-section provenance,
transcripts never rewritten) still stand and are unchanged.
+**Cluster sharing was DEAD for 14.5% of the corpus — the half it was written for
+(2026-07-29).** `digestBatch` looked the plan up with `${channelSlug}/${directoryName}`; the
+plan is keyed `${channelSlug}/${metadataId}`. Those differ for **11,175 of 77,106 videos**,
+and not evenly: **7,870 are `the-quartering-rumble`, where dir !== id for every single
+video** — precisely the mirror set `digestSharing.ts`'s header cites as the win. A miss is
+silent and looks like success (the mirror is simply not recognised, so it regenerates), which
+is why it survived a bake-off, a pilot and a 102-video validation run. `videoDirForSlug` had
+the same bug from the other side. Fixed by recording `DuplicateVideoRef.videoDir` — which the
+detector already knew and threw away — only where it differs from `id`.
+
+**A cost lever is only real in AUDIO-HOURS (2026-07-29).** See the table at the top. The
+general lesson, and it is the same one the validation run taught about throughput: **the
+corpus's video count and its audio-hours are differently distributed, so any estimate that
+counts videos is measuring the wrong thing.** `bin/digest-plan.ts` exists so the next lever
+gets priced before it gets built.
+
+**`NEAR_THRESHOLD_DEFAULT` is 0.35 (2026-07-29), and it is a PUBLISHING change.** Bracketed
+at 0.6/0.45/0.35/0.25 on identical inputs; every step down is a strict superset. Marginal
+bands were read, not counted: of the 4,264 clusters admitted at 0.45, 96.2% have
+byte-identical titles, 99.5% are cross-platform, 3 are same-channel. The recall knee is at
+**0.45** — take that if a more conservative assertion is ever wanted. 0.35 was chosen because
+its band is still clean and the report's job is to surface real mirrors to readers. It was
+changed in its own commit because it changes what every built site asserts, NOT because it
+saves sweep time (it saves 1.2 days of 80).
+
+**`updateDuplicateOverride` had ZERO CALLERS before this session.** It shipped complete —
+the `confirmed` flag, the read-modify-write, the atomic rename, and a careful rule that an
+empty patch clears a decision without deleting a confirmation — and nothing in the repo ever
+invoked it. Because `clusterMaySharePartial` fails CLOSED, that meant ~166 `needsReview`
+clusters could never be confirmed by anyone, and the review queue asked a question with no
+way to answer. `shareClusterFromCanonical` was in the same state. Both are now wired from
+/actionable. **Worth generalising: "the extension point exists" is not the same as "the
+feature works", and a fails-closed gate with no way to open it is indistinguishable from a
+missing feature.**
+
+**Total digest failures persisted NOTHING, and that was the review queue's data source
+(2026-07-29).** Both total-failure paths in `digestVideo` skipped `writeDigestSection` for a
+good reason — an empty section with current provenance reads as fresh and the video is never
+retried — and the consequence was that the worst outputs left evidence only in a job log that
+rotates. `DigestRecord.failures` is a SIBLING of `sections`, so `isSectionFresh` cannot see
+it and the retry is preserved. It records `no-output` vs `all-rejected` separately, because
+"the model proposed nothing" and "the model proposed 13 chapters and a guard clamped them
+all" look identical from outside and need opposite fixes.
+
+**Writing the test for that found a second bug, of a shape this repo has now hit twice.**
+`loadDigest` rebuilds the record field by field rather than spreading, so `failures`
+round-tripped to nothing — the write succeeded and only the read omitted it. Structurally
+identical to Phase 2's `pageHashes: [""]`, and caught the same way: by asserting on what came
+BACK, not on what went in. Anything added to `DigestRecord` must be added to that reader.
+
**The viewer mode is `?vm=digest`, NOT `?vm=summary` (2026-07-29).** "Summary" was already
taken twice — `DisplaySummary` is a video *listing card* (`common/lib/transcripts.ts`) and
`/summaries/` is the browse-index page tree the search index serves. PLAN.md's own naming
