feat: Feishu model-picker card, multi-key provider groups, i18n card copy

[daydayup123v]

What does this PR do?

Two related improvements to model switching, developed together.

1. Feishu: native interactive /model card. Picking a provider, model, and reasoning effort (minimal/low/medium/high/xhigh/max/ultra, plus none and keep) now happens in one native card, and the gateway commits model + effort in a single atomic config write. Every card action re-validates the requesting operator, chat, message, bound session, and profile; stale card revisions are fenced, so a delayed click can never apply to a session that has been replaced by /new or to a chat the card does not belong to. While an agent is running, the switch is refused with the existing busy message unless the user confirms "continue" in the card.

Previously, Feishu users switched models through text commands, and there was no way to pick reasoning effort from the chat.

2. Same-URL, different-key custom providers stay separate. Two custom providers that share one base URL but authenticate with different API keys are now tracked as distinct groups through config entries, the provider catalog, slug resolution, the /model listing, session and global restore, model-discovery write-back, and the runtime credential pool — so the second key group is never hidden by, overwritten by, or restored as the first. Each group is listed under the name the user gave it when adding it; only exactly identical names fall back to a suffix.

The card copy is served from agent/i18n (gateway.model.picker.*, added to all 17 locales); the language is captured when the card is sent so an open card never repaints mid-flow. The same change fixes doubled \n escapes in the previously shipped /fast and /reasoning picker strings (15 locales), which rendered a literal backslash-n; a regression test now renders every locale's picker strings and asserts real line breaks.

Type of Change

• ✨ New feature (Feishu picker card; multi-key provider groups)
• 🐛 Bug fix (picker newline escapes)

Changes Made

• plugins/platforms/feishu/model_picker.py (new): card state machine (provider → model → effort → apply/confirm), per-card update lock, revision fencing, TTL.
• gateway/slash_commands_model.py: picker plumbing; authorization/session re-checks on every action; one atomic commit per apply.
• hermes_cli/model_switch.py, utils.py: model+effort persistence goes through one targeted read-modify-write (_CONFIG_RW_LOCK + dotted-key update) so concurrent writers are not clobbered.
• hermes_cli/custom_provider_identity.py (new) + providers.py, config_providers.py, model_switch_providers.py, main_provider_setup.py, model_setup_flows_custom.py, runtime_provider_custom.py, cli_model_switch_mixin.py, agent/credential_pool.py: one shared identity for same-URL key groups across listing, selection, persistence, restore, and the runtime pool.
• locales/*.yaml (17 files): gateway.model.picker.* keys; newline-escape fix.
• Tests: tests/gateway/test_feishu_model_picker.py (new), tests/hermes_cli/test_same_relay_key_routing.py (new), plus aligned suites; tests/agent/test_i18n.py regression for real newlines.

How to Test

1. Add two custom providers with the same base URL and different API keys; send /model in a Feishu chat — both groups appear separately, named as you added them.
2. Open the card: provider → model → effort (e.g. ultra) → Apply. The config updates model + reasoning_effort in one atomic write.
3. Press Apply twice, or after /new: the delayed click is refused with "this card's session or authorization changed"; nothing lands on the new session.

Local test evidence

• Repository test runner (scripts/run_tests.sh) across all affected areas —
  300 files: 2,367 tests passed, 13 skipped, 0 failing tests. (The only red rows
  in the sweep were two PM-toolchain suites that require the uv binary, which
  was absent from the test box; with uv on PATH they pass 11/11.)
• Focused set while developing: 48 files / 576 tests, 0 failures.

Checklist

• Conventional commit message
• Tests added/updated for the new behavior
• Exercised against the Feishu adapter on Ubuntu (WSL2)
• Full pytest tests/ sweep — CI will run the full matrix; local run covered the affected areas above