feat(gate): --supersedes — a correction points at the request it replaces

[nao-amj]

What

--supersedes <id> on gate request and gate fast-track: a forward-only link from a correcting request to the older one it corrects.

gate fast-track --supersedes 2026-04-15-0007 \
  --action "the judge was run; the earlier conclusion does not hold" \
  --reason "..."

gate show renders supersedes: above the action, and gate chain <old-id> now resolves forward to the correction.

Why

Principle 04 has described this shape since it was written — "a correction is a new record that references the old, which preserves the trail of thinking, not just the latest conclusion." The ctx passage shipped ctx supersede; gate never had the equivalent, so corrections lived in action prose.

Measured against a 2213-record content_root:

	count	share
requests using correction wording	445	20.1%
requests naming another id in prose	148	6.7%
both — a correction whose target is written but not traversable	67	3.0%

Those 67 are principle 17's failure mode: a restatement bound to nothing. This PR binds them to structure going forward. It does not backfill.

The population matters: an earlier reading of this put the number at 407. Re-measuring with a real YAML parse (rather than a grep that matched each file's own id: line) and hand-classifying a deterministic 15-record sample showed most correction-worded records are not correcting a gate request at all. The mechanism's job is to bind the 67, not to invent corrections.

Design decisions

A flag, not a verb. ctx could afford a verb because it has one record-creating shape. gate has two entry points, and a correction-only verb would leave the other unable to correct anything. A flag on an existing verb also inherits its principle 13 shape instead of requiring a new classification.

Forward-only. The correcting record carries the link. The superseded record is never mutated and there is no inverse superseded_by — writing one would mean amending an immutable record. Both stay in the ledger.

Validated before allocation. A target that does not resolve is refused at the write boundary before the new id is allocated, so a bad link never leaves a half-written record behind. Self-supersession is rejected in the domain.

chain learns the field. Corrections used to name the older id in action text, which gatherRequestText picked up for free. Moving that mention into a structured field would otherwise have made chain blind to exactly the links the field exists to make traversable — more precise and less reachable at the same time. source_agora_play has that gap today; this one is closed at introduction.

Not included

• Any taxonomy of correction kinds — --reason carries which shape applies (claim wrong / grounds wrong but judgment stands / addendum / withdrawal).
• Surfacing corrections from gate show <old-id>. The inverse lookup costs a full scan: 0.11s versus 0.91s on the same root, an 8.3x difference. It belongs on chain, which now resolves it.

Verification

• npm test: 1954 → 1962 (+8). Failing set identical to the branch point in both directions (26 pre-existing on main).
• KNOWN_FLAGS ↔ schema.input.properties drift detector green; the doc-sync guard required and received the docs/plugin-schema.md row.
• Exercised end to end against a live content_root: the record renders, and gate chain <old-id> traverses to it.
• Example and fixture ids are synthetic, per 5195a5f.

🤖 Generated with Claude Code

https://claude.ai/code/session_01HbrpT13G92kRawJXZuzH49