fix: modernize review executor and verdict semantics

[kosako]

Summary

• make personal-codex-review a capability-checked, ephemeral read-only executor using native review --base/--commit/--uncommitted modes
• fail closed on author/routing, CLI capability, target OID, invalid ref/option boundaries, or clean-worktree mismatches without mutating the checkout
• keep GitHub lifecycle and draft-before-diff authorization in personal-review-request
• move severity/verdict/independence semantics into a self-owned output contract
• add positive and adversarial eval specs, including mixed/unknown authors, wrong PR targets, invalid refs, and unique stacked eval IDs

Closes #214

Independent review

• Routing preflight: all five commits are authored by Codex, so the required reviewer is Claude.
• Restacked Fable review (4f91e2a..d2a3999): must 0 / should 1 / nit 3, verdict WARNING.
• First fix/re-review (4f91e2a..b13eb9f): earlier four findings RESOLVED; must 0 / should 1 / nit 2, verdict WARNING.
• Final cumulative Fable re-review (4f91e2a..50dc8e6): all seven prior findings RESOLVED; must 0 / should 0 / nit 1, verdict APPROVE.
• Remaining nit is optional wording alignment in one adversarial eval; it does not weaken the fail-closed behavior.

Validation

• all 17 scripts/tests/*.sh self-tests
• scripts/check-manifests.sh — 21 manifests passed
• scripts/check-injection.sh — only expected pre-existing findings in unrelated assets
• scripts/build.sh — 42 artifacts built
• scripts/register.sh — 42 registered, pending 0
• current Codex CLI accepts native review modes, stdin prompt, read-only sandbox, config override, ephemeral mode, --base=<ref>, and --commit=<oid>
• Git ref validation and rev-parse --end-of-options behavior checked
• eval IDs checked unique across changed eval files
• staged public-safety gate passed