Goal

Give deslop a structured input so it spends LLM tokens only where slop is likely. Two outputs:

1. query slop-fixes — pinpoint structured fix actions (Haiku-tier: confirm shape, apply). AST/graph signals only, no embedder needed.
2. query slop-targets — ranked file/module targets with tier field (sonnet=file-level, opus=cross-file/module). Uses embeddings if present, degrades gracefully to AST/graph signals.

New crate: analyzer-embed

Separate Cargo crate in this workspace, separate binary agent-analyzer-embed. Heavy deps (ort, tokenizers) isolated here; main agent-analyzer binary stays small.

Two model variants, picked at install time (skill prompts user):

Released as 5 platforms × 2 sizes = 10 release assets. Model file downloaded + cached next to binary, content-hashed.

Granularity picked at install time:

• compact: per-file × 128 dim
• balanced: per-function × 256 dim
• maximum: per-function × 768 dim

Storage: sidecar .claude/repo-intel.embeddings.bin (packed int8/fp16). Main repo-intel.json stays diffable.

Subcommands on agent-analyzer-embed:

• scan — full re-embed of all files, JSON to stdout
• update — delta-only: read existing sidecar, hash files, re-embed changed/added, drop removed
• version

New subcommands on main agent-analyzer binary

• set-embeddings --input - — accept JSON via stdin, merge into sidecar (mirrors existing set-descriptors / set-summary pattern)

New queries

query slop-fixes

Returns structured fix actions for Haiku to apply:

```json
{
"fixes": [
{"action": "delete-file", "path": "debug.log", "reason": "tracked log artifact"},
{"action": "delete-lines", "path": "src/auth.ts", "lines": [142, 158], "reason": "orphan export legacyHash — 0 importers"},
{"action": "delete-lines", "path": "src/api.ts", "lines": [88, 90], "reason": "empty catch block"},
{"action": "remove-dep", "manifest": "package.json", "name": "lodash.merge", "reason": "no import resolves"},
{"action": "replace-lines", "path": "tests/auth.test.ts", "lines": [12, 14], "with": "", "reason": "tautological assertion"}
]
}
```

Categories (all AST/graph, no embedder needed):

• Orphan exports (symbol exported, 0 importers in graph)
• Orphan files (not imported, not entry-point, not test)
• Tracked artifacts (.log, .bak, .orig, .DS_Store, coverage/, dist/ in git)
• Unused deps (declared in manifest, no import resolves)
• Empty catch blocks (AST shape)
• Tautological tests (AST: expect(x).toBe(x), assert!(true))
• Orphan snapshots/fixtures (file exists, no test references it)
• Duplicate constants (same literal/string defined N+ places)
• Long-skipped tests (it.skip / #[ignore] + git age >180d)
• Old TODOs (TODO/FIXME/XXX/HACK regex + git blame age >180d)
• Old @ts-ignore / #[allow] / eslint-disable (regex + age)
• Stale CI configs (.travis.yml when .github/workflows/ exists)
• Two-of-same tooling (eslint + biome, prettier + biome, multiple lockfiles)

query slop-targets

Returns ranked targets for Sonnet/Opus models:

```json
{
"targets": [
{"path": "src/worker.ts", "tier": "sonnet", "score": 8.7, "suspect": "defensive-cargo-cult", "why": "hotspot 2 + bugspot 5 + 1.6x comment density"},
{"area": "src/auth/", "tier": "opus", "score": 9.1, "suspect": "over-abstraction", "why": "4-deep wrapper chain, single impl per layer"}
]
}
```

Sonnet tier uses file-level signals (combine hotspot + bugspot + size anomaly + comment density + bot authorship + recent big-drop).

Opus tier requires new graph traversals (see below).

Opus-tier graph traversals (v1 scope)

New work in analyzer-graph:

• Single-impl chains: trait/interface with exactly one concrete implementor across the import graph
• Wrapper towers: chains where every node has fan-in 1 + fan-out 1 (A→B→C all pass-through)
• Duplicate logic: AST-shape isomorphism + (if embedder present) embedding similarity for semantic duplicates
• Cliché-name clusters: helper / utility / manager etc. flagged when clustered, not just present

NLP-enabled patterns (require embedder)

Layered on when embedder is installed; degrade silently when not:

• Comment-restates-code: embedding similarity between comment text and next N tokens of code > threshold
• History-prose detection in docs: classifier on past-tense + change-reference patterns
• Stylometry-based AI authorship: per-file embedding distance from repo's own human-authored baseline (replaces removed metadata-based aiAttribution)
• Semantic duplicates: function-level embedding similarity (catches what AST-shape match misses)
• Doc-drift v2: README prose section ↔ function semantics similarity

Schema additions (backward compat)

All new fields use #[serde(skip_serializing_if = \"Option::is_none\", default)]:

```rust
pub struct RepoIntelData {
// ... existing fields
#[serde(skip_serializing_if = "Option::is_none", default)]
pub embeddings_meta: Option, // model id, dim, granularity
// raw vectors live in sidecar .embeddings.bin
}
```

Why not bundle the embedder

Default install stays ~10MB. Power users opt into ~30MB or ~195MB explicitly. Architectural symmetry with main binary download flow: skill prompts → JS wrapper downloads → cached locally.

Acceptance criteria

• analyzer-embed crate compiles, produces standalone binary
• Both model variants (small/big) load + embed successfully
• agent-analyzer-embed scan and update produce well-formed JSON
• agent-analyzer set-embeddings merges JSON into sidecar
• Sidecar persists across init / update runs
• query slop-fixes returns valid fix actions across all categories above
• query slop-targets returns ranked targets with tier and why
• Opus-tier graph traversals: single-impl chains, wrapper towers, duplicate logic
• All new fields backward-compatible (load existing v0.5.0 artifacts without error)
• Tests for: empty embeddings (no embedder), partial embeddings (delta update), full embeddings
• CI builds 10 release assets (5 platforms × 2 sizes)

Companion work

Skill orchestration + UI prompts: agent-sh/repo-intel#TBD