fix(discovery): stop blanket-excluding OpenRouter from serving on ZDR grounds

[seonghobae]

Why

Reported (in the org's product-loop session): OpenRouter contributes zero servable free models even though OPENROUTER_API_KEY is registered and OpenRouter's own catalog carries real, complete per-token pricing (unlike Bytez, which never reports pricing at all — separate, already-tracked gap in #940).

Root cause: PROVIDER_MODEL_SOURCES's openrouter entry has carried evidence_only=True since 952996ec ("fix(discovery): keep OpenRouter catalog evidence-only"). evidence_only=True means no OpenRouter model can ever become a serving agent — agent_from_discovered hard-refuses (raise ValueError("evidence-only model cannot become a serving agent")) and is_routable_discovered_model excludes it — regardless of that specific model's own ZDR status.

Traced the actual security history through the commit trail (721ae519, 588aef3, 984fe91, 6dcfc07, caab2ed, 952996ec):

• The real, well-documented bug 721ae519 fixed ("fix(security): scope ZDR evidence to OpenRouter") was cross-provider evidence contamination: an earlier version matched OpenRouter's ZDR feed against every discovered provider's models by model-id suffix, so a different provider's model that merely shared a display name with an OpenRouter model could incorrectly inherit OpenRouter's ZDR attestation. That was correctly fixed by requiring an exact canonical id match — no more suffix matching. This fix is real, correct, and untouched by this PR.
• 952996ec (empty commit body, no linked issue/design doc) is a separate, later, much broader change: it doesn't touch matching logic at all — it just adds evidence_only=True to OpenRouter's source, which bars the entire provider account from ever serving. There's no independent security rationale on record for this specific scope.
• Worse, _apply_discovered_model_evidence's zdr_capable=not model.evidence_only and matches(model.model_id) meant OpenRouter's own rows — even ones OpenRouter's own exact-match ZDR feed attests as ZDR-capable — could never actually carry that evidence into a serving decision, since evidence_only=True zeroed it out before matches() even mattered.

ContextualWisdomLab/.github's own scripts/ci/zdr_policy.py (used by the central Strix/Noema/OpenCode review sidecar) was already built to consume this correctly: is_zdr_model() evaluates OpenRouter per exact "provider/model" route against the same /api/v1/endpoints/zdr feed, with full PROVIDER_ZDR_SCOPE/PROVIDER_CREDENTIAL_NAMES/PROVIDER_BASE_URLS entries for openrouter already in place. It never depended on this blanket exclusion — it was just never reachable, because contextual_orchestrator's own discovery layer never let an OpenRouter model become a servable candidate in the first place.

ZDR is per-route/model evidence — this repo already models it that way for every other provider (zdr_capable, supports_zero_data_retention, privacy:zdr/privacy:no_zdr tags) — not grounds to bar an entire provider account from serving.

What changed

• Removed evidence_only=True from OpenRouter's PROVIDER_MODEL_SOURCES entry. No other source ever set it; is_routable_discovered_model/agent_from_discovered's evidence_only guard remains in place as general infrastructure for a genuinely evidence-only future source.
• Updated the discover_all_models comment describing OpenRouter's role.
• Updated 3 existing tests that asserted the old blanket-exclusion behavior (test_default_sources_request_openrouter_full_modality_catalog, test_enable_cheapest_activates_the_lowest_priced_discovered_agent, test_enable_cheapest_bootstraps_independent_provider_accounts) to assert the corrected behavior — OpenRouter now legitimately wins cheapest-price selection when it is in fact cheapest.
• New regression test test_openrouter_free_zdr_verified_model_is_routable_and_servable: proves a free, ZDR-attested OpenRouter model is routable, serving-agent-eligible, and carries privacy:zdr — using the real production PROVIDER_MODEL_SOURCES entry, not a synthetic fixture.

Verification

• RED-before-GREEN: stashed the production fix, confirmed the new regression test fails exactly on assert openrouter_source.evidence_only is False (i.e. it's True on unfixed code); restored the fix, confirmed it passes.
• python -m pytest tests -q (deselecting the same pre-existing, unrelated fast_mlsirm environment gap noted in fix(discovery): bound OpenRouter free-model endpoint fan-out to one deadline #939): 2824 passed, 1 skipped.
• interrogate contextual_orchestrator/model_discovery.py: 100%.

Scope note

Left _apply_discovered_model_evidence's not model.evidence_only guard in place — it's correct in the abstract (an evidence-only row can never serve regardless of its zdr_capable value, so gating zdr_capable computation on it is immaterial dead weight for such rows, not wrong), and now correctly inert for OpenRouter since evidence_only is False there.

---

Generated by Claude Code

[coderabbitai]

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

• Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[seonghobae]

Superseded by #949 ("fix(discovery): route OpenRouter by model evidence", merged 8cd99f1), opened concurrently by another session. It removes the same evidence_only=True from OpenRouter's PROVIDER_MODEL_SOURCES entry — and goes further, adding apply_openrouter_spend_admission() to fail closed on paid OpenRouter routes without a verified positive credit balance (free routes and every other provider stay unconditionally admitted). That's a case this PR didn't cover: my fix only proved a free ZDR-attested OpenRouter model becomes servable, without addressing what should happen to paid OpenRouter rows once the provider is no longer blanket-excluded.

Closing without merging to avoid duplicating already-landed, more complete work.

---

Generated by Claude Code