feat(serder): #79 pluggable serialization backend — seam + DirectJson zero-copy writer

[joeldsouzax]

Summary

Implements the #79 design (reviewed in #138): serialization becomes a pluggable component behind the EventSerializer seam, with serde_json as the unchanged default and a hand-rolled DirectJson backend selectable per call via serialize_with. Zero new dependencies.

The seam (commit 1)

SAID/version orchestration is written once in serialize_with: render once with a placeholder SAID + zero-size version string, backpatch the measured size in place, hash, splice the SAID into the EventLayout slots (both fixed-width). Backends only control how bytes are produced. The old per-ilk pipelines refactor behind the seam with unchanged public signatures; as a side effect even the serde_json reference drops from 3 full JSON renders per event to 1 render + 2 in-place patches.

DirectJson backend (commit 2)

Writes the five fixed event grammars straight into the caller's buffer — field names/framing as compile-time constants, values through an RFC 8259 escaper, slot offsets recorded while writing. No Value tree, no intermediate String.

Also fixes a panic found while mirroring the threshold renderer: tholder_to_json hit 0/0 integer division on a (0, 0) weight, and parse_weight accepted "0/0" from the wire — so a crafted event with a valid SAID deserialized fine and panicked on re-serialization. parse_weight now rejects zero denominators; weight rendering is division-free and shared by both backends. Bug probes red-before/green-after.

Conformance gates (commits 2–3)

• Cross-backend proptests per ilk: DirectJson output must be byte-identical to SerdeJson over generated events (boundary sn/bt/thresholds, all five seal variants, both identifier derivations)
• Escaper oracle: property-tested against serde_json's own string encoder
• Cross-path: direct output SAID-verifies through the unchanged read path
• Allocation budget (tests/serder_allocation.rs): counting allocator asserts direct < serde_json allocations
• Existing round-trip / kel-chain / keripy differential suites pass untouched (1872 tests)

Benchmarks (benches/serder.rs, CodSpeed-wired)

Local aarch64-darwin, both figures include the fixed Blake3 hashing cost:

bench	serde_json	direct
icp	4.08 µs	2.32 µs	1.76× faster
ixn + 16 anchors	7.33 µs	4.20 µs	1.75× faster

Breaking change (0.x minor)

New SerderError::InvalidEventLayout variant (backend-reported layout inconsistencies surface as typed errors, never a corrupt frame). Everything else is additive: EventSerializer, EventRef, EventLayout, SerdeJson, DirectJson, serialize_with.

Follow-ups (per design §3.3/§3.5)

• Strict canonical read-path parser with offset-based SAID verification (card to be filed; feeds C-a C-a · Zero-copy event input — borrow-ify KeriEvent across serder #129)
• Default-flip + serde demotion to dev-deps after the direct backend soaks

Closes #79

🤖 Generated with Claude Code

[codspeed-hq]

Merging this PR will not alter performance

✅ 21 untouched benchmarks
🆕 4 new benchmarks
⏩ 21 skipped benchmarks¹

Performance Changes

	Benchmark	BASE	HEAD	Efficiency
🆕	icp_direct	N/A	24.7 µs	N/A
🆕	icp_serde_json	N/A	47.6 µs	N/A
🆕	ixn16_direct	N/A	36.5 µs	N/A
🆕	ixn16_serde_json	N/A	75.1 µs	N/A

---

_{Comparing feat/79-serialization-backend-seam (2747123) with main (c6f1448)}

Footnotes

1. 21 benchmarks were skipped, so the baseline results were used instead. If they were deleted from the codebase, click here and archive them to remove them from the performance reports. ↩