feat(gcp-to-aws): AI clarify detect-confirm cards + AI-only mini assumption sheet

[herosjourney]

Summary

The problem, in plain English: The plugin's code scanner usually already knows which AI framework and models a startup uses — but the moment detection is anything less than fully certain, the questions fall back to cold walls of options: a 7-option framework taxonomy and a 19-option model list (Gemini 3.5 Flash Thinking vs Gemini 3.1 Flash-Lite vs GPT-5.4 Nano…). A founder who just watched the plugin correctly find langchain in their requirements.txt is then asked to pick their framework from scratch, as if the scan never happened. And in the AI-only flow there's no assumption sheet at all — the full Clarify wizard shows users what was detected before asking anything, but AI-only users go straight into question batches.

Changes:

• Q14 partial-detection confirm card (clarify-ai.md): when detection produced signals but not full resolution, present "I detected LangChain (requirements.txt) — that's right / edit / just LangChain" instead of the full taxonomy. The full option list only appears cold when detection found nothing.
• Q19 confirm card + family-first cold path (clarify-ai.md): detected-but-uncertain models are offered as a short pick-list of the actual candidates. When nothing was detected, ask provider family first (Gemini / GPT / o-series / other), then show only the matching subset — never the 19-option list in one message. The full list remains the interpretation catalog.
• AI-only mini assumption sheet (clarify-ai-only.md, new Step 1.75): detected values (framework, model, capabilities) and to-be-assumed defaults presented as a compact confirm-or-edit sheet before Batch 1, mirroring the full wizard's sheet-before-questions contract. Sheet-resolved questions are skipped in the batches.
• Q27 separated (clarify-ai.md): Activate-credits question asked as its own short follow-up after the technical essentials ("Last one, and it's about money rather than tech"), never mixed into the technical batch. Existing anti-inference rules (no eligibility from spend, no silent default) untouched.

Multi-workload confirmation already had accept/edit/drop actions — unchanged.

Why

Usability review: detect-confirm is the natural completion of the extraction machinery this skill already has (chosen_by: "extracted", confidence scores, auto-skip rules). The only missing piece was what to do at partial confidence, which is the common case.

Test plan

• mise run lint:md — 0 errors
• Q14/Q19 interpretation rules, override hierarchy, and Bedrock baseline tables untouched (diff-verified — wording additions only above the option lists)
• AI-only fast path (Step 1.5) unaffected; mini sheet only runs on the standard/complex path
• Manual: profile with models[0].confidence = 0.6 produces a two-candidate confirm card, not the full catalog

Out of scope

• Bedrock mapping tables / design-ai changes (explicitly none — model recommendations unchanged)
• Full Clarify wizard batching (PR feat(gcp-to-aws): Clarify interaction contract — menu actions, plain-language corrections, batched essentials #170)
• Multi-workload merge/split actions (v2 per existing limitation note)

[herosjourney]

Pushed 731cdba (after merging current main, which now includes #170): the batch-composition line #170 added to clarify.md still listed Q27 inside the AI/agentic essentials batch, contradicting this PR's Q27 separate-follow-up rule in clarify-ai.md. The batch list now excludes Q27 and points at the presentation rule, so an agent reading both files gets one consistent instruction. Lint green (855 files).