feat(cli): adjudicate --apply interactive merge walk (#137, Slice 2b-ii)

[jasonssdev]

Part of the #139→#137 curation-output arc (Slice 2b-ii, final slice). Advances #137.

What

Gives adjudicate a path to action: --apply walks the SAME verdicts and, per eligible group, previews what will fuse then prompts before merging. Reuses the 2b-i prepare_merge/merge_core blocks verbatim — the destructive core is untouched.

Per eligible group: stale-id pre-check → prepare_merge preview → Merge <absorbed> into <survivor>? [y/N/skip] → merge_core → per-merge autocommit.

Guardrails (destructive, but reversible via unmerge)

• Eligible = verdict SAME AND exactly 2 members. DIFFERENT/UNCERTAIN never merge. N>2 SAME groups → skipped (N>2, merge manually).
• Survivor = member_ids[0] (alphabetical-first), absorbed = member_ids[1].
• Prompt: only y/yes applies; empty/n/skip/other decline.
• Stale-id guard: a member absorbed by an earlier merge this run → skipped (member already merged), never crashes.
• Mid-run failure (prepare_merge or merge_core) stops the run (exit 1); prior per-merge commits remain and are unmerge-reversible.
• --apply + --json rejected (exit 2, before any read); --apply + --same-only is a no-op.
• End summary: applied X, skipped Y (N>2 / already-merged / declined).

Scope

• Additive: non---apply adjudicate (plain / --json / --same-only) is byte-identical. Only adjudicate + tests touched; merge/unmerge/prepare_merge/merge_core unchanged (diff is +142/-0 in main.py).
• Deferred (still out): guarded unattended batch (--apply-same, no confirm), gated on bug: adjudicate reports a flat 0.95 confidence and marks part-whole pairs as SAME #138 verdict quality; N>2 merging; survivor heuristics beyond alphabetical.

Verification

• Full suite: 2023 passed (2024 with the added coverage test); ruff/format/mypy clean.
• Strict TDD, 19 --apply tests incl. accept→merge→ledger, unmerge round-trip, stale-id guard (overlapping groups), each prompt answer, mid-run failure (both prepare_merge and merge_core), --apply --json rejection, summary counts.
• Independent verify: PASS WITH WARNINGS (0 critical). Bounded risk review: 0 data-loss findings — survivor/absorbed direction confirmed correct (member_ids = tuple(sorted(...))), preview==apply (same prepared object), partial-failure recoverable, destructive core untouched. The one coverage WARNING (prepare_merge failure branch) is now closed by an added test.