fix(providers): honor target_model on the custom-provider resolve path

[liuhao1024]

What does this PR do?

Makes resolve_runtime_provider() honor a caller's explicit target_model on the named-custom-provider path, closing a precedence gap against every other branch in the same function.

Today the custom branch (_resolve_named_custom_runtime) fills the returned runtime's model key from the provider-level custom_providers.<name>.model default — on both the credential-pool and non-pool paths — and resolve_runtime_provider() returns it as-is, never consulting target_model. Every other branch already applies target_model or model_cfg.get("default") precedence (generic explicit path, Nous, opencode, Azure Foundry), so the custom branch is the one place where a per-call model silently loses to the provider default.

Concrete impact (#93092): an auxiliary.background_review task configured with provider: custom:<name> + model: <X> resolves to the provider's default model instead of X. _resolve_review_runtime() then marks the fork routed=True and replays a compact digest instead of full history — so the review runs on the parent's model with digest quality loss, paying the routed-path cost for zero benefit, with no warning anywhere.

The fix applies the same precedence at the custom branch return: when target_model is set, it overrides the runtime's model. A target_model of None keeps today's behavior exactly.

Related Issue

Fixes #93092

Type of Change

• 🐛 Bug fix (non-breaking change that fixes an issue)

Changes Made

• hermes_cli/runtime_provider.py — in resolve_runtime_provider()'s named-custom branch, set custom_runtime["model"] = target_model before returning when target_model is provided, mirroring the target_model or default precedence used by the other branches. Single point covers all of _resolve_named_custom_runtime()'s internal return paths (bare-custom direct alias, credential pool, plain custom entry).
• tests/hermes_cli/test_runtime_provider_resolution.py — three regression tests: target_model outranks the provider default; the provider default stands when no target_model is passed; target_model lands in the runtime even when the provider declares no top-level model.

How to Test

pytest tests/hermes_cli/test_runtime_provider_resolution.py -q — should pass (64 passed), including the three new tests.

• test_target_model_outranks_custom_provider_default_model: provider entry with model: hermes, resolved with target_model="Curador" → rt["model"] == "Curador". Observed result: on unpatched main this fails (rt["model"] == "hermes"); with this PR it passes.
• test_custom_provider_default_model_stands_without_target_model: same provider entry, no target_model → rt["model"] == "hermes" (today's behavior preserved; passes on main too by design).
• test_target_model_sets_model_when_provider_has_no_default: provider entry without a top-level model, resolved with target_model="Curador" → rt["model"] == "Curador". Observed result: on unpatched main the key is absent; with this PR it carries the caller's intent.

Checklist

• Code follows the project's style guidelines
• Self-review completed
• Comments added for complex logic (precedence rationale in the branch comment)
• Tests added that prove the fix works
• New and existing unit tests pass locally (64 passed in the resolution suite)
• Platform: macOS (logic is platform-independent provider resolution)

[alt-glitch]

This was generated by AI during triage.

Duplicate of #78573: both repair named custom-provider target_model precedence over the provider default, including pooled and direct resolution. #78573 is the earlier open implementation with regression coverage.

[liuhao1024]

The Python tests / e2e failure here is a known main flake, not caused by this change:

• Failing test: tests/e2e/test_platform_commands.py::TestSlashCommands::test_plaintext_restart_gateway_in_group_stays_plain_text[telegram] — send.assert_called_once() → "Called 0 times".
• The same test with the same assertion failed two days ago on PR fix(models): keep reasoning available under metadata-only overrides #91620 (run 32656038080's sibling, run id 32497603622, 2026-08-21), an unrelated change; test(e2e): name the background-task crash behind a missing send #92130 is an open PR specifically chasing this missing-send flake, and the same failure signature appears in the checks of several other open PRs (fix(kanban): redact the card body in kanban_create before persisting #92366, fix(state): constrain last_active to a sane epoch window #91543, ...).
• This PR's diff is confined to resolve_runtime_provider()'s custom-provider branch (the runtime dict's model key) plus new unit tests in tests/hermes_cli/test_runtime_provider_resolution.py — no overlap with the Telegram slash-command/restart/send path that test exercises. Full tests/hermes_cli/test_runtime_provider_resolution.py passes locally: 64 passed.

[teknium1]

Closing as duplicate with credit — @web-wyf's #78573 implemented the same fix three weeks earlier and was salvaged via #93356. Your test coverage confirmed the same root cause independently; the landed version covers both pool and direct paths. Thanks.