fix(slackbotv2): forward reasoning efforts for custom-provider models

[badjer]

Problem

reasoningForModel() in services/slackbotv2/src/console-session-link.ts validates a requested reasoning effort against a hardcoded table of known Codex model ids, and returns undefined for any model it does not recognize. Deployments that point Codex at an OpenAI-compatible endpoint through a custom model_providers.* entry — the pattern supported since #1410 and described in the production docs with CODEX_MODEL / CODEX_MODEL_PROVIDER — necessarily use model names that table can never contain.

The effect is that a user's explicit -rsn high (and any SLACKBOTV2_CHANNEL_DEFAULTS reasoning entry) is silently discarded on such deployments: no error, no feedback, and every turn runs at the deployment default effort.

Fix

Models absent from the per-model table are now validated against the canonical effort set (the union of every effort any Codex model accepts, including minimal for the Nanocodex mapping) and forwarded. The provider is the authority on what its own models support and rejects unsupported efforts itself — a visible upstream error is strictly better than a silently ignored request. Known models keep their exact per-model validation, including the gpt-5.6 alias and dated-snapshot handling.

This deliberately changes one documented expectation: reasoningForModel('codex', 'gpt-5.3', 'high') previously returned undefined (unknown-model-as-typo). A mistyped model fails the turn anyway with a model-not-found from the provider, so pre-rejecting its effort adds no safety — while the old behavior broke a real, docs-recommended configuration.

Tests

• New: unknown models forward every canonical effort (codex and nanocodex, including minimal); non-canonical values (superduper, empty) still return undefined; non-Codex harnesses unchanged.
• Updated: the gpt-5.3 expectation moved into the forward-through case per the rationale above.
• Rebased on current main; ran the CI-equivalent locally (pnpm install --frozen-lockfile, then bun test test in services/slackbotv2): 251 pass / 1 skip / 0 fail.
• check:types fails identically on a clean main checkout (fetch.preconnect typings in test/slack-user.test.ts, untouched by this change, under Bun 1.3.x); CI does not gate slackbotv2 on check:types.

🤖 Generated with Claude Code