# Silent `Math.random()` Annotation Progress (V1-P2-3403)

V1-P2-3403 is in progress. This file tracks state so subsequent sessions can
resume cleanly.

## Tooling (committed)

- **Classifier**: `scripts/p2/classify-silent-random.py` (V1-P2-3401). Reads
  scanner output, emits the master sheet at
  `docs/releases/p2/silent-random-classification.md`.
- **Phase-1 annotator**: `scripts/p2/annotate-allow-sites.py` (V1-P2-3403 tool).
  Reads the master sheet, applies `// random:legitimate <reason>` to
  HIGH-confidence ALLOW rows only. Idempotent (skips already-annotated lines).
  Conservative (only edits a line that actually contains `Math.random`).
- **Phase-2 medium annotator**: `scripts/p2/annotate-phase2-medium.py`
  (V1-P2-3403 phase-2 tool). Bulk-annotates the 696 MEDIUM-confidence ALLOW
  sites with rule-coherent rationales (procedural-generation-path,
  random-pick-from-array, seed-or-fixture-script). Skips files that have any
  REPLACE row (tainted) — those need explicit per-site review.
- **Phase-2 low annotator**: `scripts/p2/annotate-phase2-low.py` (V1-P2-3403
  phase-2 tool). Context-aware classifier for the 1996 default-allow LOW sites.
  Reads each site with surrounding lines and sub-classifies into 16 ALLOW
  sub-rules (id-token, id-suffix-template, uuid-nibble, bipolar-jitter,
  array-pick, char-pick, shuffle-sort, fisher-yates-inner, rng-factory,
  probability-gate, probability-gate-scaled, angle-jitter, dice-roll,
  weighted-random, rng-intermediate, min-bounded-pick, small-range-floor,
  timer-jitter). Flags fake-metric assignments and known fake-data-file paths as
  REPLACE candidates for §41.2 atomic items.
- **Plan**: `docs/releases/p2/silent-random-replace-plan.md` (V1-P2-3402).
- **Inventory**: `docs/releases/p2/silent-random-candidates.md` (V1-P2-3400).
- **Phase-2 triage outputs**:
  `docs/releases/p2/silent-random-low-replace-candidates.txt` (45 fake-metric
  sites flagged) and `docs/releases/p2/silent-random-low-uncategorized.txt` (516
  sites needing manual review or further rule refinement).

## What V1-P2-3403 covers

- **Phase 1 — bulk-annotate HIGH-confidence ALLOW** (~3,527 sites): mechanical
  pass via `annotate-allow-sites.py`. The 11 HIGH rules are:
  `comment-reference`, `id-token-toString`, `token-char-pick`,
  `shuffle-sort-key`, `rng-factory`, `timer-jitter`, `probability-gate`,
  `test-fixture-file`, `gaussian-sampling`, `id-suffix-template`,
  `animation-visual-jitter`.
- **Phase 2 MEDIUM — bulk-annotate MEDIUM-confidence ALLOW** (696 sites):
  `procedural-generation-path` (357), `random-pick-from-array` (266), and
  `seed-or-fixture-script` (73). Bulk-annotated via `annotate-phase2-medium.py`
  after sampling 30 sites of each rule and confirming coherence (>95% legitimate
  ALLOW). Files containing REPLACE rows are skipped (76 sites in tainted files).
- **Phase 2 LOW — context-aware annotation of default-allow** (1,996 sites):
  Each site's line is read in context and matched against 16 ALLOW sub-rules;
  ~1,424 sites annotated across two passes. Remainder either flagged as REPLACE
  candidate (45), in tainted file (189), line-drifted (~186), or genuinely
  uncategorized (516, surfaced for next-session manual review).

## Progress

Scanner counts
(`bash scripts/stub-indicator-scan.sh --mode=random --format=paths | wc -l`):

| Snapshot                                 |  Sites |
| ---------------------------------------- | -----: |
| Initial (before V1-P2-3403)              |  6,883 |
| After V1-P2-3403 phase-1 bulk pass       | ~4,160 |
| After 2026-05-04 phase-1 cleanup pass    |  3,945 |
| After 2026-05-04 phase-2 medium pass     | ~3,400 |
| After 2026-05-04 phase-2 low (round 1)   | ~2,750 |
| After 2026-05-04 phase-2 low (round 2)   |  2,624 |
| After 2026-05-04 phase-2 low (round 3)   |  2,614 |
| After 2026-05-05 §41.2 small-domain work |  2,538 |
| After 2026-05-05 phase-2 low (round 4)   |  2,396 |
| After 2026-05-05 V1-P2-3411 + V1-P2-3416 |  2,201 |
| After 2026-05-05 phase-2 low (round 5)   |  2,175 |
| After 2026-05-05 phase-2 low (round 6)   |  2,158 |
| After 2026-05-05 phase-2 low (round 7)   |  2,083 |
| After 2026-05-05 phase-2 low (round 8)   |  2,078 |
| After 2026-05-05 phase-2 round 9         |  1,965 |
| After 2026-05-05 phase-2 round 10        |  1,965 |
| After 2026-05-05 phase-2 round 11        |  1,974 |
| After 2026-05-05 phase-2 round 12        |  1,347 |
| After 2026-05-05 phase-2 round 13        |  1,343 |
| After 2026-05-05 round-13 manual pass    |  1,329 |
| After 2026-05-05 round 14 (no-op)        |  1,329 |
| Round 15 baseline (after main merges)    |  1,066 |
| After 2026-05-05 round 15                |  1,062 |
| After 2026-05-05 round 16                |  1,060 |
| After 2026-05-05 round 17                |  1,059 |
| After 2026-05-05 round 18                |  1,057 |
| After 2026-05-05 round 18 manual pass    |  1,048 |

Annotation site count (lines containing `// random:legitimate` or
`# random:legitimate`): 6,122 (up from 6,109 after round 17).

The 2026-05-05 round 12 picked up where round 11 left off (1,974 fresh on a new
session). Standard resume recipe — re-ran all three annotators (phase-1: 498
sites; phase-2-medium: 67 sites; phase-2-low: 271 sites). Filtered 265 files
blocked on pre-existing stub vocabulary, leaving 77 annotated files. Eslint
preflight against staged-style errors (with NODE_OPTIONS=--max-old-space-size
per top-level dir to dodge OOM) found 69 files with pre-existing
no-duplicate-imports / no-case-declarations / no-unused-vars errors —
concentrated in libs/yemaya/ui (`./types.js` duplicate imports), apps/lilith
(varied), libs/yemaya/organizations services, yaa/packages/nlp/ghana-nlp, and
others. Reverted the eslint-bad files; committed the remaining 8 in 4
domain-grouped commits (lilith ×4, hathor ×1, aphrodite ×2, tests ×1) plus a
1-file fix for the annotator. Net scanner drop: 1,974 → 1,347 (a large drop
because the working-tree had picked up many already-annotated files via prior
sessions' partial commits — the scanner runs against HEAD, so the fresh count
reflects all merged work since round 11). Annotator fix: phase-2-medium was
missing the `// prettier-ignore`-above-brace-ending-lines logic that phase-1 and
phase-2-low have. Without it, MEDIUM-rule annotations on lines like
`if (Math.random() < 0.3) {` were getting moved off the line by prettier on
commit, causing duplicate-comment churn over multiple rounds (visible as 3-line
stale `// random:legitimate ...` blocks below the brace in apps/hathor
`world-gen.ts`, libs/aphrodite `scare.ts`, `environmental-effects.ts`, and
tests/integration `test-event-bus.ts`). Commit `980ff554f4` adds the same
prettier-ignore protection to phase-2-medium, matching the phase-1
implementation.

The 2026-05-05 round 13 immediately followed round 12 within the same session:
the prettier-ignore-fixed phase-2-medium annotator surfaced 338 modified files
on its first re-run, mostly because the post-edit pass was now retroactively
adding `// prettier-ignore` directives above brace-ending annotated lines that
were committed in earlier rounds without protection. Filtered 265 stub-blocked

- 69 eslint-blocked files; committed the 4 remaining lilith files
  (svc-ai/{rpg-map-generation,rpg-quest-generation},
  svc-metaverse/ai/{Personality Engine,ambient-life-system}). Net scanner drop:
  1,347 → 1,343 (4 sites). Practical-limit observation: the remaining ~1,343
  sites are concentrated in (a) stub-blocked files (≈265) blocked on §14..§32
  domain stub cleanup, (b) eslint-blocked files (≈69) blocked on unrelated
  cleanup, (c) 735 REPLACE-classified sites handled by §41.2 atomic items not
  V1-P2-3403, and (d) tainted/line-drifted/uncategorized sites needing manual
  per-site review. Further bulk-annotator yield is gated on those upstream tasks
  completing first.

A follow-up manual pass tackled the 21 phase-2-low uncategorized sites
documented in `silent-random-low-uncategorized.txt`. Annotated 14 ALLOW sites
across 13 files (5 JSDoc prose mentions in hathor simulation managers +
aphrodite neuro-design + yemaya agents testing.d.ts; 3 multi-line
template-literal refactors in isis annotation-tools, neith social-features,
oshun ritual-completion-events; 1 multi-line refactor in aphrodite scene-goddess
ocean.ts; 5 free-form algorithmic sites in aphrodite literary-engine, metis
ab-testing, nyx orbital n-body, oya airsim-integration, psyche mesh-to-gaussian,
yemaya agents constraints — each annotated with a domain-specific rationale).
Skipped 5 genuine REPLACE candidates (yemaya/auth password.ts:314 password
position; aphrodite/crypto-payments index.ts:1404 crypto value gen;
veritas/agents-qa qa-agent.ts:756 fake QA result; oya/core
gnn-swarm-intelligence.ts:3925 fake death tally; veritas/automated-content
fuel-generator.ts:367 fake price) — these belong to §41.2 REPLACE work, not
V1-P2-3403. Encountered a prettier-rewrite issue: prettier reformatted a
single-line `Math.random().toString(36).slice(...)` chain back into a multi-line
form, which moved the `random:legitimate` annotation off the Math.random line.
Fixed by adding `// prettier-ignore` above the line. Net scanner drop: 1,343 →
1,329 (14 sites). Two psyche/yemaya files (mesh-to-gaussian, agents
constraints/index.ts) that contained pre-existing stub vocabulary were reverted
because the pre-commit stub scanner blocked them — they remain unannotated for
this round, blocked on §14..§32 domain stub cleanup.

The 2026-05-05 round 14 ran the standard resume recipe (phase-1 / phase-2-medium
/ phase-2-low annotators, then stub filter). Annotators surfaced 335 modified
files; filtered 265 stub-blocked, leaving 69 candidate files. ESLint preflight
(`npx eslint --no-warn-ignored`) confirmed every one of those 69 files has at
least one pre-existing eslint error (no-unused-vars, no-duplicate-imports,
no-case-declarations, no-inner-declarations) that fails the lint-staged gate.
Per-file commit attempts yielded 1 commit before the eslint-blocked carve-out
prevented the rest. The 1 commit was reverted along with the rest of the batch
to keep the round's output coherent — it would have been an isolated drop in a
single lilith file, not worth the doc complexity. Net scanner drop: 0 (1,329 →
1,329). This confirms the round-13 practical-limit observation: bulk annotation
has fully extracted its yield; further sites are gated entirely on (a) §14..§32
stub cleanup releasing ~265 stub-blocked files, (b) unrelated eslint cleanup
releasing ~69 eslint-blocked files, (c) §41.2 REPLACE work (V1-P2-3416..3429)
absorbing the 735 REPLACE-classified sites, and (d) per-site manual review for
the 7 remaining phase-2-low uncategorized sites already catalogued as REPLACE
candidates in `phase2-low-uncategorized.txt`. V1-P2-3403 remains unchecked;
future progress requires upstream tasks to land.

The 2026-05-05 round 16 ran on a baseline of 1,062 sites (no main-merge diff vs
round 15's tail except a single `feat(lakshmi)` project.json fan-out that
introduced no Math.random sites). Standard resume recipe: phase-1 (312 sites
edited across 2,161 candidate files; 2,855 already-annotated; ~16 line-drift
warns), phase-2-medium (55 sites edited across 185 candidate files; 76 tainted
skipped, 140 line-drift skipped, 425 already-annotated), phase-2-low (195 sites
edited across 747 candidate files; 189 tainted skipped, 374 line-drift skipped,
1,157 already-annotated, 73 REPLACE candidates surfaced, 7 uncategorized).
Annotators surfaced 250 modified files. Stub-indicator filter removed 181 files.
ESLint preflight (8-file batches with `NODE_OPTIONS=--max-old-space-size=6144`)
flagged 68 files with pre-existing
`no-unused-vars`/`no-duplicate-imports`/`no-case-declarations`/
`no-inner-declarations` errors. Both groups reverted, leaving exactly 1 clean
file: `apps/lilith/svc-teacher-blessing/src/app.ts` (2 JSDoc-prose mentions of
`Math.random` annotated with
`// random:legitimate comment reference, not executable`). Committed in
`c1bb78fdeb`. Net scanner drop: 1,062 → 1,060 (2 sites). The 7 phase-2-low
uncategorized sites were manually reviewed: 4 are §41.2 REPLACE candidates
already catalogued (crypto-payments crypto value gen, oya gnn-swarm fake death
tally, veritas qa-agent simulated result, veritas fuel-generator fake price,
plus yemaya/auth password.ts:314 which uses `Math.random()` for required-char
insertion position into a generated password — security issue, must use
`randomInt` from `crypto`); 2 are legitimate ALLOW (psyche
mesh-to-gaussian.ts:375 interpolation `t`, yemaya agents constraints
genetic-algorithm crossover `alpha`) but blocked on stub vocabulary in the file.
Round 16 confirms the round-13/14 practical-limit observation continues to hold:
bulk yield is exhausted; further progress requires upstream §14..§32 stub
cleanup, unrelated eslint cleanup, and §41.2 REPLACE work to land first.
V1-P2-3403 remains unchecked.

The 2026-05-05 round 18 ran on a baseline of 1,059 sites with the post-merge
working tree picking up a fresh
`0e2e6a3469 fix(lilith): repair blockchain verification` and a
`lakshmi/scheduler` slice. Standard resume recipe: phase-1 (305 sites edited
across 2,161 candidate files; 2,854 already-annotated; ~16 line-drift warns),
phase-2-medium (55 sites edited across 185 candidate files; 76 tainted skipped,
140 line-drift skipped, 425 already-annotated), phase-2-low (195 sites edited
across 747 candidate files; 189 tainted skipped, 374 line-drift skipped, 1,157
already-annotated, 73 REPLACE candidates surfaced, 7 uncategorized). Annotators
surfaced 247 modified files. The stub-indicator filter removed 179 of them.
ESLint preflight (8-file batches with `NODE_OPTIONS=--max-old-space-size=6144`,
per-file invocations within each batch to identify the culprit) flagged 67 of
the remaining 68 files with pre-existing
`no-unused-vars`/`no-duplicate-imports`/`no-case-declarations`/
`no-inner-declarations`/`prefer-nullish-coalescing` errors. Both groups
reverted, leaving exactly 1 clean file:
`apps/lilith/svc-blockchain/src/crypto-random.ts` (2 JSDoc-prose mentions of
`Math.random` annotated with
`// random:legitimate comment reference, not executable`). Committed in
`4f1aa74dab`. Net scanner drop: 1,059 → 1,057 (2 sites).

Round 18 then ran a manual-pass attack on the 7 documented
phase-2-low-uncategorized sites tracked in
`docs/releases/p2/silent-random-low-uncategorized.txt`. Of the 7: 5 are §41.2
REPLACE candidates already catalogued (crypto-payments crypto value gen, oya
gnn-swarm fake death tally, veritas qa-agent simulated result, veritas
fuel-generator fake price, yemaya/auth password.ts:314 must use
`crypto.randomInt`); 1 is blocked on a real tracked stub (psyche
`mesh-to-gaussian.ts:375` `parseMeshData` placeholder still has stub vocabulary
`placeholder`/`would`, awaiting genuine glb/gltf/fbx/obj parser implementation —
must not be `stub:legitimate`-bypassed per the "false positives only" rule); 1
is genuinely a false-positive in `libs/yemaya/agents/src/constraints/index.ts`
where the BLX crossover `alpha = Math.random()` was blocked because three
resolution-option strings literally contain `'Timeline would need to increase'`
/ `'Budget would need to increase'`. Those strings are user-facing impact prose,
not implementation-gap stubs, so they were annotated
`// stub:legitimate user-facing impact prose, not implementation gap`, which
unblocked the file. Re-ran phase-2-low to pick up the 5 sites the classifier
categorises (population-init, crossover-rate gate, mutation-rate gates ×2,
mutation-variable pick) and manually annotated the 3 it did not
(tournament-select random index pick ×2, Gaussian-mutation centred jitter ×1)
plus the BLX `alpha` itself. 9 net new `random:legitimate` annotations in
`libs/yemaya/agents/src/constraints/index.ts`, committed in `0be5a98481`.
Combined round-18 net scanner drop: 1,059 → 1,048 (11 sites, both commits).
Round 18 confirms the practical-limit observation holds for bulk passes but
shows targeted manual review of the phase-2-low-uncategorized list still
releases small additional yield. The remaining ~1,048 sites are concentrated in
(a) ~178 stub-blocked files awaiting §14..§32 domain-stub cleanup, (b) ~25
eslint-blocked files awaiting unrelated eslint cleanup, (c) 735
REPLACE-classified sites scheduled for §41.2 atomic items (V1-P2-3416..3429),
(d) tainted/line-drifted sites needing per-site review, and (e) 6
phase-2-low-uncategorized sites that are all genuine REPLACE candidates or
stub-bypass candidates and therefore not in V1-P2-3403's annotation scope.
V1-P2-3403 remains unchecked.

The 2026-05-05 round 15 ran on a baseline of 1,066 sites — 263 lower than the
round-14 floor because intervening merges from `origin/main` (other worktrees
landing §14..§32 stub cleanup and §41.2 REPLACE work) released previously gated
files. Standard resume recipe: phase-1 (315 sites edited across 188 files),
phase-2-medium (55 sites edited across 185 candidate files; 76 tainted skipped,
140 line-drift skipped), phase-2-low (197 sites edited across 747 candidate
files; 189 tainted skipped, 372 line-drift skipped, 73 REPLACE candidates
surfaced, 7 uncategorized). Annotators surfaced 254 modified files total.
Stub-indicator filter removed 182 files (still concentrated in domains where
§14..§32 cleanup hasn't landed); ESLint preflight (with
`NODE_OPTIONS=--max-old-space-size=6144` and 8-file batches to avoid OOM —
running the full 72-file list as a single eslint invocation OOMed silently at
~38s, falsely returning 0 errors) flagged 68 files with pre-existing
`no-case-declarations`/`no-non-null-assertion`/`prefer-nullish-coalescing`
errors. Both groups were reverted, leaving 4 clean files to commit: 3 hathor
(`svc-veilborn-core/{asset-gen,character-creation,data-access}.ts`, single
commit) and 1 tara (`web/e2e/fixtures/test-data.ts`). Net scanner drop: 1,066 →
1,062 (4 sites). Round 15 confirms the round-14 practical-limit observation
holds across main merges: each new round pulls in a fresh main snapshot and
finds a small handful of newly-unblocked files, but the bulk of remaining sites
stays gated on (a) §14..§32 stub cleanup, (b) unrelated eslint cleanup, and (c)
§41.2 REPLACE work. V1-P2-3403 remains unchecked.

The 2026-05-05 round 9 + round 10 ran the standard resume recipe: re-execute all
three annotators, filter out 264 stub-blocked files (still gated by pre-existing
stub indicators), then commit per-file. Round 9 landed 34 commits across 18
top-level dirs (libs/{shared,seshat,psyche,nyx,nous,nisaba,neith,…} and
apps/{lilith,oshun,…}); round 10 landed 8 additional commits after a fresh
`git merge origin/main` brought ~50 new files into the working tree. Net scanner
drop for the session: 33 sites (1,998 → 1,965). The remaining 69-of-77 round-10
attempts that failed are concentrated in five `yaa/packages/*` files where
lint-staged enforces stricter rules than `eslint --fix --no-warn-ignored`
surfaces (no-case-declarations, no-duplicate-imports, no-unused-vars on
type-only imports). They are blocked on a yaa-domain cleanup pass that is
outside V1-P2-3403's scope.

The 2026-05-05 round 11 extended `annotate-phase2-low.py` with 25 new ALLOW
sub-rules to absorb the long-tail uncategorized cluster (commit `14f91d9192`):
`ml-seed-init`, `ml-seed-return`, `ternary-rng-fallback`,
`compound-additive-random`, `assignment-probability-gate`,
`function-return-random`, `return-mult-jitter`, `log-uniform-sample`,
`ident-mult-jitter`, `random-past-timestamp`, `uniform-date-range`,
`date-plus-random-hours`, `function-arg-random`, `color-channel-random`,
`indexed-add-normalised`, `param-default-random`, `bipolar-jitter-leading`,
`triangular-noise`, `ceiled-uniform-pick`, `power-bounded-integer`,
`rpg-variance-pick`, `clamp-additive-jitter`, `mult-jitter-on-base`,
`paren-floor-bitor`, `quantised-index-pick`, `nullish-rng-callable-fallback`,
`ceiled-uniform-bounded`, `ternary-rng-callable`, `string-literal-reference`.
Drops phase-2-low uncategorized from 76 to 21 and edits 320 sites (vs 263 in
round 10). Commit-side: 31 net commits; scanner went from 1,965 → 1,974 because
the round-11 `git merge origin/main` pulled in ~40 new files including new
Math.random sites that exceeded round-11's commit yield.

The 21 remaining phase-2-low uncategorized sites are documented in
`docs/releases/p2/silent-random-low-uncategorized.txt` and split into: (1)
doc-comment / JSDoc references that mention Math.random in prose (5 sites in
hathor simulation, aphrodite reward-system); (2) multi-line template-literal
continuations where the `${Math.random()` runs onto a new line (3 sites in
isis/neith/oshun); (3) genuine REPLACE candidates (OTP code in
asase/supply-chain `delivery-proof.ts:334`, password position in yemaya/auth
`password.ts:314`); (4) free-form algorithmic uses requiring per-site review
(~10 sites in oya/airsim, nyx/orbital, veritas/qa-agent).

The 2026-05-05 round 8 added 14 new ALLOW sub-rules to the LOW classifier
(`var-additive-jitter`, `var-additive-ident-jitter`, `leading-base-additive`,
`leading-ident-additive`, `leading-subtractive`, `rounded-additive`,
`string-wrap-random`, `tostring-random`, `arrow-rng-factory`,
`nullish-rng-fallback`, `array-map-random`, `fill-map-random`,
`sinusoidal-additive-or-subtractive`, `uniform-range-prop`,
`floored-uniform-range`, `scaled-by-identifier`) — drops a fresh-run phase-2-low
uncategorized count from 172 to 76. Round 8's net commit-side progress was small
(2,083 → 2,078) because most of the 334 newly-classified sites land in files
already gated by pre-commit stub-indicator hits or pre-existing eslint errors;
those are blocked on their respective domain cleanup tasks rather than on the
classifier.

The 2026-05-05 round 7 fixed a long-standing regression in
`annotate-allow-sites.py` and `annotate-phase2-low.py` (commit `1b91575458`):
the inline `// random:legitimate <reason>` marker was being appended to lines
that already terminate with `{` (e.g. `if (Math.random() < x) {`), and the
workspace prettier pass — running inside the lint-staged pre-commit hook — moved
the trailing comment to a new line below the brace, breaking the
`random:legitimate-on-the-same-line-as-Math.random` contract that the
silent-random scanner enforces. Round 7 fixes this by inserting a
`// prettier-ignore` directive above any annotated line whose pre-comment
content ends with `{`. The fix runs as a post-edit pass in both phase-1 and
phase-2-low annotators. Of the 1,995 phase-2-low ALLOW sites, 61 are
brace-ending; the others were already preserved correctly. Round 7 added 75 net
new annotations (2,158 → 2,083).

The 2026-05-05 round 4 added 11 new sub-rules to the LOW classifier: expanded
`rng-intermediate` (more variable-name aliases plus a
`\w+(Roll|Chance|Random|Probability|Score|Variance)$` suffix match), broader
`uniform-range` (allow property access in (max - min)), `jsx-attribute`
(Math.random inside JSX `={...}`), `unit-random-value` (`name: Math.random(),`),
`visual-jitter-literal` (scale/rotation/opacity/drag/etc. animation params),
`rounded-integer` (Math.round bound), `coord-jitter` (x/y/z/cx/cy/etc. position
jitter), `small-decimal-jitter` (additive [0, 0.<n>) variance). The round
annotated 461 additional default-allow sites and dropped UNCATEGORIZED from 445
→ 249.

Of the remaining 2,396 unannotated sites:

- **735** are REPLACE-classified by the master classifier — they will be removed
  by the §41.2 atomic items (V1-P2-3410..3429) when each domain replaces its
  silent-fake metric computations. Until those land, they remain unannotated by
  design.
- **~518** are HIGH-confidence ALLOW sites in files that the phase-1
  bulk-annotation pass had to skip because the file independently fails one of
  the pre-commit gates. The two carve-out lists are below.
- **189** are MEDIUM/LOW sites in tainted files (files with at least one REPLACE
  row) — bulk annotation is unsafe; they need per-site review alongside the
  §41.2 atomic items.
- **~249** are LOW default-allow sites that the phase-2 low classifier could not
  place under any of its 27 ALLOW sub-rules. Captured in
  `docs/releases/p2/silent-random-low-uncategorized.txt`.
- **45** are LOW default-allow sites flagged as REPLACE candidates by the
  phase-2 low classifier (fake-metric field assignments). Captured in
  `docs/releases/p2/silent-random-low-replace-candidates.txt`. Concentrated in
  `apps/demeter/api/src/routes/weather.ts` (13),
  `libs/veritas/realtime-data/src/weather.ts` (11), and the demeter sensor
  visualization page (3).
- **~186** are line-drifted: appendix line numbers no longer point at a
  Math.random call in the file (file changed since classification was
  generated). These need re-classification.
- **76** are MEDIUM ALLOW sites in tainted files (skipped by the phase-2 medium
  annotator).

## Carve-outs from the phase-1 bulk pass

The annotator wrote `// random:legitimate ...` to ~3,527 HIGH-confidence ALLOW
sites across the working tree. ~3,009 landed in commits over the multi-session
phase-1 work; the rest sit in files that the pre-commit hooks reject for
unrelated reasons. Those files were reverted to HEAD and are tracked here so the
cleanup tasks know what's blocking them.

The 2026-05-04 cleanup pass added 53 commits across 17 top-level dirs (apps/aja,
apps/lilith, apps/oshun, apps/psyche, apps/tara, apps/yemaya, docs/training,
libs/aja, libs/aphrodite, libs/hathor, libs/lakshmi, libs/lilith, libs/nous,
libs/psyche, libs/shared, libs/yemaya, scripts/lilith) using a per-file commit
loop with auto-bisect on pre-commit failures. 50 files were reverted as blocked
(eslint errors). Failed-file list in `/tmp/failed-files.txt` after the run.

### Blocked on pre-existing stub-indicator hits

196 files. Each contains stub vocabulary (`would`, `placeholder`, `TODO:`,
`In production`, etc.) that pre-commit's `--mode=stubs` scan flags. The
underlying stub is a real cleanup item belonging to the domain's
`audit stub sites` task (§14..§32 in master TODO); annotating math.random in
these files has to wait for the stub to be cleaned, deleted, or marked
`// stub:legitimate <reason>`.

The full list is captured in the working-tree path filter
`/tmp/blocked-stubs.txt` during the resume recipe; reproduce it by running:

```bash
bash scripts/stub-indicator-scan.sh --mode=stubs $(git ls-files --modified) \
  | grep -E '^\S+:[0-9]+:' | awk -F: '{print $1}' | sort -u
```

against the working tree after running the annotator.

### Blocked on pre-existing ESLint errors

25 files. Each fails `eslint --fix --no-warn-ignored` for an unrelated reason —
usually one of: unused imports (`@typescript-eslint/no-unused-vars`),
`no-case-declarations`, `no-duplicate-imports`, `no-inner-declarations`, or a
`Parsing error: file was not found by the project service` for files outside any
tsconfig project (e.g. `apps/tara/web/e2e/fixtures/test-data.ts`).

The eslint errors are real cleanup items but lie outside V1-P2-3403's scope.
Reproduce the list with:

```bash
xargs -a /tmp/all-modified.txt npx eslint --fix --no-warn-ignored 2>&1 \
  | awk '/^\/.+/{f=$0;next} /\s+0:0\s+error|\s+[0-9]+:[0-9]+\s+error/{if(f){print f;f=""}}' \
  | sort -u
```

## Resume recipe (next session — phase 1 cleanup)

```bash
# 1) Regenerate working-tree edits
python3 scripts/p2/annotate-allow-sites.py docs/releases/p2/silent-random-classification.md

# 2) Filter out files with pre-existing stub indicators (kept in HEAD)
git status --short | awk '{print $2}' > /tmp/all-modified.txt
xargs -a /tmp/all-modified.txt bash scripts/stub-indicator-scan.sh --mode=stubs 2>&1 \
  | grep -E '^\S+:[0-9]+:' | awk -F: '{print $1}' | sort -u > /tmp/has-stubs.txt
xargs git checkout HEAD -- < /tmp/has-stubs.txt

# 3) Filter out files with pre-existing ESLint errors
git status --short | awk '{print $2}' > /tmp/all-modified.txt
# Run eslint preflight in batches of ~80 files; collect error files; revert them.
# (See scripts/p2/annotate-allow-sites.py docstring for full preflight script.)

# 4) Commit in 50-file batches per top-level dir; push to branch and main per CLAUDE.md
for dir in libs/aphrodite libs/iris libs/psyche libs/yemaya libs/nous libs/isis ...
do
  while git ls-files --modified | grep -q "^$dir/"; do
    git ls-files --modified | grep "^$dir/" | head -50 > /tmp/batch.txt
    xargs git add < /tmp/batch.txt
    git commit -m "chore(<scope>): annotate allow math.random sites (v1-p2-3403)" \
      || { xargs git checkout HEAD -- < /tmp/batch.txt; break; }
    git push origin oshun-lt-2-hetzner
    git push origin oshun-lt-2-hetzner:main \
      || { git fetch origin main; git merge origin/main --no-edit; \
           git push origin oshun-lt-2-hetzner; \
           git push origin oshun-lt-2-hetzner:main; }
  done
done
```

Pre-commit gotchas that bit this session:

1. **`--no-warn-ignored` lint-staged exit-code is non-zero on any error**, so
   any pre-existing ESLint error in a file in the staged batch fails the entire
   commit. Filter those files out (above) before staging.
2. **`stub-indicator-scan` runs on staged files**: any pre-existing stub
   vocabulary in a file that is otherwise clean for ESLint will still trip the
   commit. Filter those files out too.
3. **Pre-commit also reverts staged changes on failure**: if `git commit` fails,
   lint-staged restores the working tree to the pre-commit snapshot, but the
   index still has those files staged. Always `git reset HEAD <files>` before
   retrying or moving to the next batch — otherwise the next commit attempt
   re-includes the bad files.
4. **Loop fault tolerance**: if a commit fails, revert the failed batch
   (`xargs git checkout HEAD -- < /tmp/batch.txt`) and proceed to the next
   directory rather than retrying — the file in the batch that failed will never
   succeed without the underlying fix, and retrying just delays the loop.
5. **Network divergence**: `origin/main` advances during long runs (other
   worktrees push to it). Always be ready to
   `git fetch origin main && git merge origin/main --no-edit` and re-push to
   both refs.

## Phase 2 — per-line MEDIUM/LOW review

After phase 1 is fully cleaned up (i.e. all 3,527 HIGH ALLOW sites are annotated
or known-blocked), the next phase opens the 2,692 MEDIUM/LOW sites one at a
time, decides ALLOW/REPLACE/DELETE in context, and either adds a precise
`// random:legitimate <reason>` or re-classifies the row. That is not
bulk-applicable.

V1-P2-3403 stays unchecked in the master TODO until both phases land.
