v4.x: interactive invocation by model name — @deepseek / @nova / @gpt5.5 (not just @claude)

[sfreudenthaler]

Feature request

Today interactive mode is triggered by @claude and runs whatever the caller's model_id is. A consumer should be able to invoke a specific model by name in a PR/issue comment — the way @claude works — e.g.:

• @deepseek review the IAM changes
• @nova summarize this PR
• @gpt5.5 is this migration safe?
• @claude … (unchanged)

The mention itself selects the model/route for that one interactive run, overriding (or in addition to) the caller's default model_id.

Why

• Lets reviewers pick the right model per question without editing workflow YAML or maintaining one caller job per model.
• Natural fit for the model diversity we already run (Claude writes, a non-Claude reviews) — a human can pull in a second opinion from another family on demand.
• The routing already exists: the orchestrator maps anthropic.* → claude, openai.* → codex-mantle, everything else → bedrock-generic. This is mostly a mention → model_id mapping layer on top of enable_mention_detection.

Sketch (for discussion, not a spec)

• A configurable alias map, e.g. @gpt5.5 → openai.gpt-5.5, @deepseek → deepseek.v3.2, @nova → amazon.nova-pro, @claude → ${default}.
• Mention detection resolves the alias, sets model_id for that run, routes as usual.
• Sensible default + allowlist so a comment can't invoke an un-provisioned/expensive model arbitrarily (cost + IAM scope — mantle openai.* is already gated by the BEDROCK_ROLE_ARN policy).
• Unknown alias → no-op or a one-line "unknown model" reply.

Targeting v4.x since it changes the interactive trigger contract.

Refs: #42, #43 (reviewer output quality — separate track).