feat(distill): cost/scale preflight before O(n²) semantic judgments (#76)

[aktasbatuhan]

Closes #76.

watchmen distill (semantic mode) judges every n*(n-1)/2 skill pair with an LLM call and gave no warning before the loop started. A 13-skill project is 78 calls; larger ones run into the hundreds, which can be a real surprise on a metered provider.

What it does

A preflight that runs before any judgment HTTP calls and asks to continue:

distill coder-pro
─────────────
13 skills → 78 pair judgments (78 model calls to qwen/qwen3.7-plus)
~217k input tokens
estimated cost ~$0.65 (input tokens only; output not counted)

Under an OAuth subscription, where per-token cost is meaningless, it drops the price and warns about quota burn instead:

13 skills → 78 pair judgments (78 model calls to claude-opus-4-8)
~217k input tokens
billed under your claude-pro subscription — no per-token charge, but this spends 78 calls against your quota

Design notes

• Headline is the call count. It's exact and meaningful regardless of billing — that's the real surprise the issue is about. The price is a secondary, conditional line.
• Dollars gated on a metered provider. config.active_provider() decides; OAuth providers (claude-pro / chatgpt) get the quota warning, not a price.
• No invented output tokens. The judge's reply length isn't knowable pre-run, so the dollar figure is an explicit input-only floor, not a fabricated total. Input tokens are a metadata-size approximation that mirrors the real _semantic_prompt (fixed scaffold + each SKILL.md excerpt capped at 8k chars, billed once per pair it appears in).
• No behaviour change for automation. Skipped for --json, non-interactive stdin, and the new --yes/-y flag.

Acceptance criteria (from #76)

• TTY semantic distill shows the preflight with pair count + ballpark cost
• --yes flag skips the prompt
• --json and non-interactive runs skip the prompt (no behaviour change)
• Estimate within ~2x of actual on a tested project — 1.05x on the real coder-pro bundle (13 skills, 78 pairs)

Testing

• 8 new tests in tests/test_skillmesh.py: pair count, metered cost, subscription has no dollar cost, input-only floor, pricing scaling, single-skill (zero pairs), and the skip/proceed/cancel + subscription-render paths.
• Full suite green: 518 passed, no regressions.

🤖 Generated with Claude Code