Expose exact packet byte costs and adopt SDK fitting - #21

Closes #5

The initial main baseline e2074b1 exposed only ceil(max(JSON,TOON UTF-8 bytes)/4) under a token-shaped compatibility field. It had no exact per-renderer byte costs and maintained its own unbounded fixed-point loop. Two failing-first tests reproduced the missing byte API in real temporary non-git SDK trackers; the Unicode/digest fixture measured 2301 JSON / 1775 TOON bytes while the byte fields were absent.

The packet now exposes exact self-inclusive bytes.json and bytes.toon, including the final renderer newline and all domain/SDK accounting fields. The compatibility estimatedTokens field remains, explicitly labeled by receipt.budgetUnit = heuristic-units: both renderings fit 4 * budget UTF-8 bytes. No model tokenizer, token upper-bound guarantee, or end-to-end cost reduction is claimed. Historical 817/body-only measurements and the real-data reporting harness now state their scope and unmeasured commands/receipts/retries.

The installed SDK already provides fitOutputEnvelope; no dependency pin changed. It owns convergence, finite work controls and skip admission, preserving caller row order and considering later smaller rows. The package supplies its inert JSON/TOON renderers and domain receipt. Per-format byte fields subtract the renderer-length difference from the SDK maximum: scalar accounting integers have the same decimal spelling in both formats, so that difference is independent of their digit widths. The SDK verifies the resulting complete envelope. Infeasible, unstable and work-limit errors retain safe typed remedies. Source completeness, first-bad/last-good semantics, exact omissions and executable restore instructions are preserved.

Companion G6 adoption was read-only; this PR changes only pm-http. Tracker: pm-http-33ck, retained in progress with the claim released for orchestrator verification.

Validation:

• node --test --test-concurrency=2 test/packet.test.ts: 4/4 passed, also executed through pinned pm test --run --progress.
• node --test --test-concurrency=2 test/http.test.ts test/commands.test.ts: 29/29 passed; includes actual host processes, SDK temp trackers, existing concurrency/crash schedules and both final renderer byte counts.
• npm run typecheck, npm run lint, npm run docstring (96 declarations) and npm run duplication (0%): passed without gate/threshold changes.
• Changelog regenerated after the last pm write; npm run changelog:check passed. Tracker history is a byte-prefix extension of main. Commit identity verified.
• Full npm run release:check: passed with isolated HOME and Git configuration; 143/143 tests, 100/100/100/100 coverage across 20 configured source files, 2/2 packed npm/Bun investigations against the published minimum SDK 2026.10.28. The minimum SDK exports the fitter. Strict tracker health also passes with isolated HOME (the normal host global pm-github installation has a version-skew notice outside this package).
• Full bun run release:check: passed serially, also 143/143 tests, 100/100/100/100 coverage, 2/2 packed npm/Bun investigations.
• Prior-head CI: Node 22 and Node 26 passed, including full npm/Bun gates; CodeQL also passed on 12b71dc. Fresh Node 22/26 CI passed on final head c86bcf7f2cec9a12ff26161f6c512611bca55d6f, including full npm/Bun gates and packed acceptance; final-head CodeQL remains pending. The required explicit CodeRabbit review completed clean on the final head; the other reviewer skips/neutral notices are not approval.

Behavioral revert proofs:
node test/revert.acceptance.ts issue5-packet issue5-packet-host issue5-skip issue5-byte-framing issue5-sdk-refusals: 5/5 passed. Each intentional revert ran in a disposable copy and failed its behavioral regression; restoring the fixed source passed again.

Follow-up c86bcf7 pins the revert harness to historical main e2074b1, so it remains reproducible after this fix lands. Production source is byte-identical to 12b71dc, which passed both local release gates. All five revert proofs, typecheck and lint passed again. A real disposable clone advanced origin/main to the fixed function: the pinned harness passed; restoring its former moving origin/main reference failed the identity check. Branch history remains a byte-prefix extension of current main; the claim is released and the changelog was regenerated after the final pm write.

Review loop:

• 1 completed full CodeRabbit round, covering e2074b1 through c86bcf7; completed review reports no actionable comments, and the completion reply confirms the full review finished. The whole body was read; no outside-diff or nitpick findings were present.
• 0 threads / 0 review comments: no fixes or refusals were needed. Every bot comment was voted; completion/quota notices were voted down and review content up.
• CodeRabbit's sandbox could not run ESLint because a config dependency was unavailable. This is a reviewer-tool limitation; local and both CI matrix lint gates passed separately.
• GitHub GraphQL rate limits blocked the supplied state/vote/watch helpers after review. REST independently confirmed the exact head, successful CodeRabbit and Node 22/26 checks, zero review comments, and all comment votes. Final-head CodeQL remained queued at the final CI handoff on 2026-11-06 16:46 UTC, about one hour after the last push. It has not executed and is not certified as passed. Node 22/26 full gates and the completed CodeRabbit review passed on the exact final head. The orchestrator must retain this pending scan as an explicit check.

No merge or release is requested.

Summary by Sourcery

Expose exact packet byte costs and use the SDK fitter to produce verified, bounded difference packets with clear heuristic-budget semantics.

New Features:

• Expose exact self-inclusive UTF-8 byte costs for JSON and TOON packet renderings, including newline and accounting fields.
• Adopt the SDK output fitter for bounded packet construction, row admission, convergence, and typed fitting refusals.

Bug Fixes:

• Ensure oversized rows are skipped while later smaller rows remain eligible and packet byte fields match the final rendered output.
• Preserve actionable typed errors for infeasible, unstable, and work-limit fitting failures.

Enhancements:

• Clarify that estimatedTokens and budget values are four-byte heuristic units rather than model-token measurements or guarantees.
• Update architecture, README, changelog, and real-data documentation to describe exact byte accounting and measurement scope.
• Expand behavioral and acceptance coverage for Unicode content, renderer framing, budget boundaries, SDK fitting behavior, and refusal handling.

Documentation:

• Document exact JSON/TOON byte costs, SDK fitting semantics, heuristic budget units, and the limits of historical cost measurements.

Tests:

• Add packet-focused tests covering exact byte accounting, row omission semantics, minimum and boundary budgets, and typed SDK refusal translation.
• Extend native-host and revert acceptance tests to verify renderer byte counts, SDK adoption, and regression behavior.

Chores:

• Regenerate the real-data report and retain tracker history and reproducibility metadata for the packet accounting change.