architecture: bert and opt detectors (second increment of #708)

[mgillr]

Stacks on #709 (contains its commits; review the second commit for the delta).

Adds two more pure-name detectors in the same pattern:

• BertFamilyDetector — BERT and RoBERTa encoder layouts; post-norm LayerNorms map onto the post_attn_norm / post_ffn_norm canonical slots (schema extended accordingly, plus embed_positions and the norm/head bias globals).
• OPTDetector — the OPT decoder layout (model.decoder.layers.N.*), distinguished from the Llama layout that architecture: canonical parameter-name mapping (first increment of #708) #709's detector already rejects.

Tests: per-family mapping with value identity, all outputs pass is_canonical, and cross-detector exclusivity (a Llama-layout dict matches neither new detector). gpt2/neox/phi remain for the third increment — they need fused-QKV splits, which are tensor ops and belong with the shape-bridge increment rather than this dependency-free module.

---

Note

Low Risk
Additive, dependency-free name remapping with unit tests only; no changes to existing merge execution paths.

Overview
Extends the cross-architecture canonical name space (#708) with two more pure-name detectors alongside the existing Llama-family mapper: BertFamilyDetector (BERT/RoBERTa encoder keys, including post-norm LayerNorms mapped to post_attn_norm / post_ffn_norm) and OPTDetector (model.decoder.layers.*, kept distinct from Llama via model.decoder. detection).

The shared schema gains post_attn_norm and post_ffn_norm roles plus globals for embed_positions, final_norm.bias, and head_lm.bias so OPT and encoder layouts can round-trip into the same layer_N.role.weight keys without tensor ops.

Tests cover BERT/OPT detect + canonicalize (value identity, is_canonical), a regression that OPT per-layer final_layer_norm.bias maps to pre_ffn_norm.bias, and cross-detector exclusivity on Llama-layout state dicts.

^{Reviewed by Cursor Bugbot for commit f3b02aa. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes using high effort and found 2 potential issues.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

Want fixes drafted automatically? Bugbot Autofix can create code changes for findings. A team admin can enable Autofix in the Cursor dashboard.

^{Reviewed by Cursor Bugbot for commit ae93097. Configure here.}