Preserve legacy tier compatibility without lint debt

[stranske]

Why

Consumers still pass the legacy tier= keyword to model selection. It must remain accepted without affecting profile-driven selection or creating Ruff debt.

Scope

Adopt existing PR #2810 as the implementation lane.

Tasks

• Preserve the ignored legacy keyword in both canonical and consumer-template registries.
• Add focused regression coverage that passing different tier values is accepted and cannot change profile-based selection.
• Restore mergeability and resolve review feedback.

Acceptance criteria

• Both registry copies accept tier= without lint failures.
• Profile selection is unchanged for varying legacy tier values.
• Targeted tests and template parity pass.

[stranske]

Acceptance criteria are already satisfied on main — closing

This issue scoped itself as "Adopt existing PR #2810 as the implementation lane", but the same fix had already merged to main in 885905e8 ("fix: retain ignored llm registry tier parameter", #2837, 2026-07-26). Checked each acceptance criterion against main at dbce5dd3 rather than against the PR:

• "Both registry copies accept tier= without lint failures." — tools/llm_registry.py:291-292 and templates/consumer-repo/tools/llm_registry.py:291-292 both carry del tier under the comment "Preserve the compatibility keyword while selection remains profile-based." del tier is what retires the Ruff unused-argument debt while keeping the keyword in the signature (tier: str | None = None).
• "Profile selection is unchanged for varying legacy tier values." — tests/tools/test_llm_registry_selection.py:84 test_configured_model_ignores_legacy_tier_when_profile_is_explicit asserts configured_model_for_provider("openai", profile="verifier-balanced", tier=...) returns model-balanced for both legacy-low and legacy-high.
• "Targeted tests and template parity pass." — the targeted test above exists on main, and the root and consumer-template llm_registry.py copies are byte-identical on main.

PR #2810 has been closed as superseded: its only remaining delta against main was a differently-worded comment plus a weaker duplicate of that test (it relied on the default profile instead of pinning one).

The keepalive-rail Actions-approval outage that previously blocked this lane is therefore moot for this issue — there is nothing left to implement.