feat: declare model lineage in the config instead of inferring it

[duckyquang]

Summary

A model's lineage can now be declared in the config instead of guessed from its id:

models:
  - gemini-2.5-flash                                              # lineage inferred, as today
  - {id: cxr-tuned-local, lineage: llama, tier: 8b, open_weights: true}

The id heuristic stays as the fallback, so every existing config behaves exactly as before. Config.models is still the plain list of ids (nothing downstream changed); the declarations land in a new Config.roster alongside it.

Two things that were previously unreachable now work: tier, and open_weights, which was hardcoded to False for every model built from a config. roster.cross_lineage_committees refuses a roster with no open-weights model, so that guard was un-satisfiable from a config file until now.

The run manifest gets config.resolved_roster: the roster as the run actually used it, each entry tagged declared or inferred. --dry-run shows the same thing inline.

Result

$ benchmaxxing run --stage overlap --manifest m.csv --config declared.json --dry-run
  models: gemini-2.5-flash, cxr-tuned-local
  lineages: {"cxr-tuned-local": "llama (declared)", "gemini-2.5-flash": "gemini (inferred)"}

ruff check . clean, pytest -q 609 passed. Tests cover both config shapes, the round trip, and my-tuned-cxr, which is exactly the id the prefix heuristic gets wrong.

Notes

Stacked: #251 -> #252 -> #253 -> this. Base is #253's branch so the diff is only the lineage work.

Closes #255

[sebasmos]

Approving. This unblocks the cross-lineage cascade (Gemini vs an open-weights model), which is the experiment worth doing. Let's aim there.

[maximinl]

Approve declared lineage (stacked on #253).

Keeping id-heuristic as fallback preserves existing configs; Config.roster + resolved_roster (declared/inferred) is the right place for tier / open_weights, which unblocks cross_lineage_committees from a config file. Merge after the stack below it.

[sebasmos]

Nice config design, but it only pays off once a real cross-lineage roster runs, @duckyquang please use the nvidia cluster #265/#95. Holding so we merge it with that result rather than ahead of it.

Superseded by the later Changes-requested review on this PR; dismissing the stale approve to avoid a mixed signal.

[Agastya191]

Good work here, and thanks for closing the loop on the lineage thread from #251: declaring lineage in the config and threading model_specs into both run_holes_test and build_committee is exactly the right seam, the roster/models split leaves every existing caller unchanged, and stamping resolved_roster with declared-vs-inferred provenance into the manifest is a genuinely nice touch. One thing I'd flag is that the bare-id path still routes through _lineage, which splits on the first separator, so an org-namespaced id like meta-llama/llama-3.1-70b still resolves to meta rather than llama. Because the override only fires when lineage is declared, a mixed-convention roster still gets a wrong within- vs cross-lineage partition in run_holes_test and the run completes clean, with lineage_source: inferred in the manifest as the only trace after the fact. It's reasonable to keep the heuristic as the documented fallback, but I'd have model_specs (or lineage_sources) warn when an inferred id looks org-prefixed, so the failure I hit in #251 is loud at run time instead of only discoverable in the manifest afterward.

[duckyquang]

@Agastya191 done in 3ed12d9. model_specs now warns when a lineage is inferred (not declared) from an id that carries a /, so the meta-llama/llama-3.1-70b → meta case is loud at run time instead of a manifest post-mortem. The heuristic stays the documented fallback and declaring lineage in the config silences it, so the bare-id happy path is unchanged.

@sebasmos understood — holding for a real cross-lineage roster run on the nvidia cluster (#265/#95) so it lands with that result rather than ahead of it. I've got the access for that one.

@maximinl thanks. Added the run-time warning for org-prefixed inferred ids (3ed12d9) so the id-heuristic fallback can't silently split a mixed roster wrong.

[sebasmos]

Approving. runner.py:116 warns when lineage is inferred from an org-prefixed id and declaring it silences the warning, which is what @Agastya191 asked for. Note the cross-lineage roster run is still open on #265, where I flagged that its cross_lineage field reads off declared ModelSpec.lineage and so could not have returned false with Gemini behind both slots. Merged main in, post-squash conflicts resolved to the branch side, suite green.