perf(sqlite): reuse canonical proof encoding in retained reads

[LIHUA919]

Refs #4224. Design and the unmodified failed qualification evidence were recorded first in #4224 (comment).

Goal and delivered outcome

The formal 1 MiB, 10k/100k SQLite D2 run on measurement source e99a83b63c149685625ecc80e5065c31b2ad36d6 failed the existing read budgets: at 100k commits, readReceipt p95 was 269.03 ms (limit 50 ms) and scanCommitted(100) p95 was 801.81 ms (limit 250 ms). The 64 KiB, 100k/300k headroom run had no failed budget rows. Both reports remain historical evidence; this PR does not relabel either result.

SQLite v2 verifies every retained row from the covering checkpoint. With a stable 1 MiB projection, each row repeatedly canonicalized the same long JSON string for the exact state and commit SHA-256 proofs. A real SQLite 128-commit probe reproduced about 302 ms for a window-end receipt and 503 ms for a 100-row scan; raw indexed SQL was under 1 ms. CPU samples concentrated in canonical byte generation.

This PR retains the exact v0 proof bytes while reusing bounded encodings of repeated long strings within one verification or archive-audit call. A later empty delta reuses its verified predecessor state digest, still compares the row's declared digest and parent lineage, and still recomputes the full per-row commit proof. The first commit always derives its own state digest. No cache crosses a read call.

Scope and compatibility

• Changed surfaces: SQLite retained receipt/scan/history verification and the internal canonical digest helper; focused TypeScript regressions.
• Persisted SQLite schema, checkpoint interval, stored state/commit digests, original receipts, writer behavior, CLI/API response shapes, provider selection and RFC thresholds are unchanged.
• Rollback: revert these read-path commits. SQLite v2 databases written before or after remain readable; no migration, re-encoding or history deletion is involved.
• Entry points: the existing AuthorityStore read/scan methods and real Python CLI are affected. No Dashboard, Lark or packaged frontend contract changes.

Verification

• Node 22.22.3: canonical digest parity across JSON key/order, sparse array, Unicode and cache eviction cases; a fixed 1 MiB payload's 100 distinct proofs reuse the encoding while matching the old digest.
• Real SQLite: large unchanged projections retain independent scan rows; forged intermediate state-digest/parent lineage fails closed; an empty root remains readable.
• SQLite TS suite: 292/296 passed in the sandbox; the four Python-consumer tests could not start their local Effect runtime there and passed 4/4 in an isolated real runtime. Migration/bounded-profile/capacity tests 19/19; real Python-to-TypeScript SQLite CLI tests 5/5; TypeScript typecheck and git diff --check passed.
• Risk-based loopx canary premerge --from-git-diff --git-diff-base origin/main passed: 3 direct checks, 5 catalog checks, 8 risk-profile checks, zero failures and zero manual holds. Its temporary local dependency link was removed before push; the tracked diff contains only the five paths above.
• Matched local 1 MiB / 128-commit probe: fixed-projection window-end receipt about 302→43 ms and scan 100 about 503→98 ms. A changing-projection receipt remains about 63 ms, above 50 ms. This probe is diagnostic, not formal D2 qualification.

A full reference-runtime matched-1m 10k/100k rerun on this exact head is still required. Retain the failed row if that rerun misses either threshold. Domain workload, stable RSS, large-history recovery, 24-hour consumer lag, restore/upgrade/rollback, OS/runtime coverage, the natural-time soak and provider promotion remain separate D2 holds. No default-provider change or live-Goal migration is requested.

Future-facing pass

The bounded encoding cache stays with the canonical codec owner and is consumed only by SQLite verification. It removes repeated proof work without introducing a second semantic owner or a speculative persisted proof format. Further proof-format work should follow measured real-workload gaps and a separate compatibility review.