feat(coordinator): add the R9 role-ablation policies (no_thinker / no_trirole)

[minion1227]

The gap

SPEC §1.3 R9 — "removing SVF / Thinker / tri-role / penultimate-token all hurt" — has a merged verifier and no producer. analysis/ablations.py (#376) takes {full, no_svf, no_thinker, no_trirole, last_token} and reports each ablation's drop. But grep those four variant names across src/, configs/ and scripts/ and they appear only inside docstrings and the example JSON at scripts/ablations_report.py:11. Nothing in the repo could ever emit one of them, so R9 could only ever be graded against numbers typed in by hand from the paper.

This PR is the producer for the two role ablations. no_svf / last_token are a different concern (parameter packing and hidden-state extraction respectively) and are deliberately left out — see Scope below.

What's here

coordinator/ablations.py adds AblatedPolicy, which satisfies the session's existing Policy protocol (decide(transcript_text, *, sample, rng) -> (agent_idx, Role)), so it drops into run_trajectory with session.py, policy.py and slm.py untouched.

variant	surviving roles
no_thinker	Worker, Verifier
no_trirole	Worker only (single role → no role differentiation)

Agent (pool-model) selection is passed through unchanged: R9's role ablations remove role structure, not model routing. There's a test asserting the ablated agent choice equals the un-ablated one on every input.

Why the mask goes on the logits, not the output

Masking is applied to the role logits before the softmax, making the result a true restricted categorical. Taking the full model's decision and remapping it afterwards is not the same thing: if the un-ablated argmax was Thinker, a post-hoc rule cannot know whether Worker or Verifier was runner-up, so it cannot reconstruct the restricted argmax. Sampling has the same problem — the restricted distribution renormalizes over the survivors rather than deleting one outcome from the full one.

That distinction is pinned by test_masked_argmax_is_not_a_post_hoc_remap_of_the_full_argmax, which constructs logits where the two rules disagree.

Scope, stated honestly

• Only the role half of R9. This produces no_thinker and no_trirole. It does not produce no_svf or last_token.
• [OUR CHOICE] no_trirole collapses to Worker. The SPEC names the ablation but not which role survives; Worker is the role whose prompt asks for a direct solution. It's a constructor argument, so the choice is overridable rather than baked in.
• This ships the producer, not the numbers. Actually filling in R9's accuracies still means running the benchmark; nothing here claims to have reproduced the paper's table.

Why it's low-risk

• Purely additive. No existing caller changes behaviour; nothing is wired into the submission or training path. The wrapper never mutates the policy it wraps (pinned by a test).
• Not a fitness-semantics change. It restricts decisions at eval time; θ keeps its full width and configure forwards it unchanged, so an ablated run stays directly comparable with the full model's.
• Reuses the shipped selection rule rather than forking it. The wrapper's argmax/sample rule is pinned against the real LinearHead.select on random hidden states (test_torch_role_ablation_parity.py), so the two cannot drift and the ablated numbers stay comparable.
• Torch-free at import. import trinity.coordinator.ablations pulls no torch; it is imported lazily inside decide(), as policy.CoordinatorPolicy.decide already does — and only after the encoder/head check, so the rejection path doesn't pull it in either.

A note on the test_no_torch_imported guards

Several modules ship an in-process test_no_torch_imported, which is order-fragile: any earlier test that imports torch fails it. test_union_oracle.py solves the same problem in a clean subprocess and is immune. Both new torch checks here follow the subprocess pattern, and they pass PYTHONPATH for the checkout under test rather than relying on whichever trinity happens to be installed.

Verification

Run against a locally-resolved main (see below), compared A/B with the same tree minus these files:

	base	with this PR
passed	1,724	1,772 (+48)
failed	11	11 — byte-identical list

The 11 are pre-existing test_report_script_* subprocess failures, untouched by this branch.

• ruff check src/ scripts/repo_governance/ → clean; mypy on the new module → clean

Note on CI: main has unresolved merge-conflict markers in src/trinity/analysis/__init__.py (introduced by dc15f97), so the package does not import and this PR is red on the merge commit. This branch touches no file in analysis/.

The fix is not contested: #402, #405, #407, #408 and #403 all rewrite that file to the byte-identical result (every one is index 9916618..6865bc2), so merging any one of them repairs main and the rest become no-ops on that file. #407 / #408 / #403 are already green — they touch that file for their own feature and resolved it in passing. Only PRs that don't touch it (this one, #415, #416, #417, #418, #412) are red.

🤖 Generated with Claude Code