fix(docs): separate interviewer Q from Mid/Principal answers, not a dialogue

[rcha0s]

Summary

• Across 32 docs, ## Interviewer probes was shaped as a two-turn dialogue: Mid: "<question>" followed by Principal: <answer>, which reads as Mid asking a question and Principal answering it. The intended shape (per CONTEXT.md) is one shared interviewer question with two independent answers at different depth.
• Fixed by promoting the embedded question to its own bold line, writing a new short Mid-level answer (correct but surface-level, missing the nuance/incident/tradeoff the Principal answer covers), and leaving every Principal answer byte-for-byte unchanged.
• 5 files in the original candidate list (37-data-and-model-poisoning.md, 39-excessive-agency.md, 56-a2a-protocol.md, 58-rag-architecture-attacks.md, 59-vector-stores.md) were checked and found already correctly shaped -- their Mid: lines use quoted short-answer phrases, not restated questions, which the initial detection grep couldn't distinguish -- so they were left untouched.
• Also fixes a quality regression the batch-fix agents introduced: 92 instances of "Yes — ...", "No — ...", "Not really — ..." constructions in the new Mid answers, violating this repo's no-em-dash style rule (a rule I missed restating in the original fix prompt). Rewritten using periods, colons, or parens per the existing convention. 3 pre-existing em-dashes outside the Interviewer probes section (Sources entries, invariant tables) were left alone as out of scope for this fix.

Test plan

• pytest tests/ -- 11/11 passed.
• mkdocs build --strict -- clean, no errors.
• Repo-wide grep confirms zero remaining instances of the broken pattern.
• Repo-wide grep confirms zero em-dashes remain in any Mid: answer line.
• Every Principal: answer verified unchanged (only the Mid: line and question formatting were touched per file).

https://claude.ai/code/session_011f1KhQCYiUUPFKco7pguLC