fix(gateway): deduplicate within a payload without requiring history (OX-M1)

[ojassug]

Audit OX-M1. DECISIONS §67. Third Lane B item.

The defect

runSessionDedupStage treated within-payload repetition as a side condition of cross-turn
matching. The dedup branch was gated on previousBlockHashes.has(item.contentHash), and the stage
returned early unless that set was non-empty — so on turn 1, three identical blocks in one payload
all survived.

The README's savings table said "the same block repeated within one payload → saves", with no
qualifier. It was true from turn two.

Why the gate was wrong, not merely conservative

recoverable: true means an intact copy survives elsewhere in the same outbound payload. Rule 3
guarantees that by preserving the first occurrence and eliding the ones after it. That claim is
verifiable from the payload alone — the model has seen the content, in this request, so the marker
resolves.

Nothing in it refers to previous turns. previousBlockHashes was answering a different question —
is this content old? — and its answer was deciding something it does not bear on.

DECISIONS §16 and §41 are untouched. They concern a sole copy elided across turns, where the
consumer is a stateless provider API with no rehydration mechanism, so the marker is deletion rather
than reference. That case is still elided with recoverable: false, still scored in full by
DriftTracker, and still fails the gate.

Measured, over real sockets

Block repeated three times in one payload:

	before	after
turn 1, block ×3	8,459 sent, 8,459 forwarded	saves
turn 2, block ×3 (already worked)	saves	saves
cross-turn sole copy	0 bytes, falls back	0 bytes, falls back
turn 1, all blocks distinct	no saving	no saving

The last row is the control that keeps the first honest — turn 1 did not become "always saves". The
third is §41 still holding, and its pin in gateway-dedup-reality.test.ts is unchanged.

The change

Two gates, relaxed by the same reasoning:

• shouldAttemptDedup is now previousBlockHashes.size > 0 || hasRepeatedContent. The occurrence
  map moved above the early return, since within-payload repetition is by itself a reason to run.
• The per-item branch fires on seenInEarlierTurn || repeatsInThisPayload.

Rule 3 and the isRecoverable = survivingHashes.has(hash) computation are unchanged, which is
what keeps sole-copy elision out of the new path: the first occurrence is always preserved, so
anything elided under repeatsInThisPayload is recoverable by construction.

What this does not establish

• Nothing about cross-turn saving. Invariant 8 stands. The ordinary conversational shape — one
  copy per payload, seen before — still saves nothing and still falls back. That is the number the
  README leads with and it has not moved.
• Nothing about the corpus. This stage never runs on the CLI or MCP paths, which execute
  plan.stageIds and don't include it. The instrument is the Gateway integration suite.
• No claim that this is a large win. It closes the gap between what the README says and what the
  code does. Whether real agent traffic repeats a block inside a single payload often enough to
  matter was not measured, and the honest framing stays the README's: use Gateway mode for
  interception, validation and metrics, not for compression.

Verification

npm run typecheck, npm run lint, npm run build and npx vitest run all pass: 89 files /
819 tests. The turn-1 case was confirmed failing against the unfixed tree first, with both
pre-existing pins passing throughout.

🤖 Generated with Claude Code