Add opt-in embedded deps surface to ir.strategy.Package (#62)

[thorwhalen]

What

Adds an opt-in embedded deps surface to ir.strategy.Package (issue #62, tracking #61). Package(embed_deps=True) emits one extra Surface(kind="deps", granularity="field") serializing the bare dependency names (version specifiers / extras / markers stripped via ir.graph._dep_name), e.g. "Depends on: sentence-transformers, networkx, meshed".

• The deps bag is its own surface, separate from prose, so a rare library name isn't diluted and the BM25 leg picks up exact dep-token matches.
• Deps remain a filter field too. Appended last, so the description (0) / readme_chunk surface_index contract is unchanged.
• Default False (progressive disclosure — today's behavior). Folds into the strategy id, so toggling re-decomposes incrementally; round-trips via strategy_to_spec. deps_template overrides the serialization.
• Touches only ir/strategy.py (deps are already scanned by sources._scan_packages).

Per the #38 decision rule this is a single-shot retrieval-quality change → lands in ir.

Why

The 231-package run found the dependency list was the single most discriminative signal, yet Package stored deps as a filter field only and never embedded them — so the index couldn't see that a package depends on meshed/networkx (⇒ graphs) or sentence-transformers/ef (⇒ embeddings).

Measured (all-MiniLM, hybrid, k=20, vs Package() baseline)

Using the #66 harness + the private benchmark:

theme	nDCG@20	distractor fp_rate@20	hard-pos recall@20
embeddings	0.861 → 0.884	0.071 → 0.071	0.40 → 0.40
graphs	0.745 → 0.745	0.222 → 0.222	0.167 → 0.167

Per-package (the real story): dep-revealing packages climb sharply — allude 83→28 (depends on meshed), xcosmo 103→40 (depends on cosmograph) — while a true non-dependent stays flat: au 22→23 (no graph dep, so deps correctly don't lift it — the FP guard holds). This confirms the lift is real dependency signal, not added text volume.

As an equal-weighted surface the lift mostly lands in the rank 20–50 band, so aggregate nDCG@20 moves modestly (+0.023 embeddings) and a few tail items jitter ±1–4 ranks. Getting allude/xcosmo into the top-20 needs deps-surface up-weighting at fusion → filed as #68. Whether to flip embed_deps on by default (and rebuild the live packages corpus) is gated on that experiment.

Tests

6 new tests/test_strategy.py cases (off-by-default, surface emitted + appended last, empty-deps → no surface, custom template, _default_deps_text strip/dedup/lowercase, spec round-trip). Full suite 420 passed; ruff E/W/F/B clean at LL88.

Part of #61. Closes #62. Follow-up: #68.

https://claude.ai/code/session_01D229oNHVN1drd1mdbQL5MV