analyze.py: dedup against existing ADRs belongs in the deterministic layer

[hailcpy]

Problem

SKILL.MD preflight says: "Note existing ADR slugs — skip any chunk whose diff is already covered by an existing ADR (match by commit hash in the Links section)." But analyze.py doesn't implement this — it only computes next_adr_number(). Dedup is left entirely to agent discipline, which is exactly the kind of step that gets skipped on re-runs, producing duplicate ADRs for the same decision with new numbers.

Re-running on a growing repo is the primary long-term usage pattern (generate once, re-run quarterly), so incremental behavior should be deterministic, not prose.

Fix

In analyze() preflight:

1. Scan the scope-resolved output_dir for NNNN-*.md files.
2. Extract linked SHAs from each (the template's ## Links → Commits: line; tolerate short SHAs via prefix match).
3. Mark any candidate whose commit set is fully covered ⇒ classification: "already_covered" (with the covering ADR's filename in signals). Partially covered ⇒ keep, but add a partially_covered_by note so the agent can extend the existing ADR instead of writing a new one.
4. Report covered_candidates count in preflight.

Acceptance

• Fixture repo with an existing ADR linking commits A,B: re-running analyze marks the A+B candidate already_covered; a new candidate C is unaffected.
• Short-SHA links (7 chars) match full SHAs.