fix(compression): price a region without registering it (OX-M5)

[ojassug]

Audit OX-M5. DECISIONS §68. Fourth Lane B item, and the last one that doesn't need a decision
from you.

The defect

trimRegionsToCeiling renders a marker for every candidate span in order to price it — which is
correct, since the marker is variable-length and self-describing and a saving estimated without it
would overstate every region. But the renderer it was handed, markerFor, also called
hasher.registerBlock(...). So every span the ceiling considered and discarded was written into
the store anyway.

target ratio	blocks registered	markers actually emitted
0.10	12	5
0.05	12	3

The store held one block per candidate, not per elision. Memory grew with how much the scanner
found rather than with how much was removed, and hasHash / expandBlockHash answered for
placeholders present in no output anywhere.

Invisible on the CLI, which supplies no hasher at all. It matters on MCP, where the server
instance is long-lived and the hasher is the reversibility store.

The constraint that governed the fix

Pricing and emission must render identical bytes, or the ceiling is computed against a different
string from the one written and --target-reduction-ratio adherence shifts.

That holds by construction rather than by care: renderElisionMarker is a pure function of the
text, its noun and its hash, and hashContent is pure. Registration never contributed to the
output. priceMarker renders the same bytes without it; registration now happens only on the two
paths that actually elide.

Corpus A/B — 578 of 578 rows byte-identical

Corpus frozen at 48ac6c8, 289 files, 578 rows, both CLI routes, ratio 0.3. Comparison engines
built with an src-only tsconfig; dist hashes 7a3bad0f515f (baseline) and 83e544289dc4
(candidate), so neither arm was compared against itself.

Zero rows differ, on any of the ten compared fields — outputSha, byteIdentical,
tokenBefore, tokenAfter, reduction, fallbackUsed, driftScore, debtScore, planMode,
stageCount.

And it is not a vacuous zero: 101 of those rows actually reduced, so markers were genuinely
priced and emitted. The corpus contains the shape the change touches — the check §56 exists to
demand.

What this does not establish

• The corpus cannot see the defect itself. It runs CLI routes, and the CLI supplies no
  TokenHasher, so the registration branch was already a no-op there. The A/B proves
  output-neutrality — the constraint — and nothing about the leak. The leak is measured by
  token-hashing-store-pollution.test.ts, which supplies a hasher directly.
• Nothing about MCP memory in practice. The counts above come from a 12-region fixture. How much
  a real long-lived session accumulated was not measured — only that growth tracked candidates
  rather than elisions.
• Nothing about bench. It supplies a hasher too and was equally affected, but its runs are
  short-lived, so the accumulation had nowhere to build up.

Verification

npm run typecheck, npm run lint, npm run build and npx vitest run all pass: 90 files /
822 tests. Both count cases were confirmed failing against the unfixed tree first, with the
byte-identity guard passing throughout — which is what isolates the leak from the constraint.

🤖 Generated with Claude Code