fix: remove quadratic router payload processing

[makosblade]

Problem

Four request-processing paths did repeated work proportional to payload size, so cost grew quadratically:

• Tool-argument repair re-serialized the whole JSON document on every edit (SJSON). 16,384 coercions took 4.75 s; 16,384 duplicate-key deletions took 181 ms.
• SSE framing re-scanned the buffer from byte 0 on every write. A 1 MiB frame arriving in 4 KiB chunks took 3.3 ms of pure scanning.
• Responses history cleanup (stripping our badge/footer before re-sending to OpenAI) rebuilt the input array per message. 1,024 messages took 117–161 ms.
• Escalation assembly rebuilt text/tool-call strings per streamed fragment.

An earlier version of this PR fixed the speed but introduced two behavior regressions (nested-union repairs lost edits; empty assistant shells stopped being dropped) and left member lookup quadratic.

Solution

• sse.Scanner: a per-buffer cursor that resumes the delimiter search where the last write left off. Zero-alloc, same LF/CRLF grammar. Wired into all 18 buffered consumers with their existing reset/EOF semantics.
• Tool arguments: an ordered, lazily-indexed edit tree. Single small edits still use the original GJSON/SJSON path; a second edit builds a per-object key index so lookups stop scanning siblings. Array wrapping preserves child edits.
• Responses cleanup: single pass, restores dropping already-empty assistant text shells.
• Escalation: retained builders; benchmarks now exercise real growing fragments instead of no-ops.

Impact

Medians of 10 runs, base vs this branch, Apple M4 Pro.

Nominal request sizes (what most traffic looks like):

Path	Input	Base	Fixed
Tool args, 1–8 fields repaired	typical tool call	1–16 µs	1–20 µs (0.8–0.9×)
Tool args, clean input, any size	no repair needed	47 ns–301 µs	+4%, still 0 allocs
Tool args, 128 fields repaired	large structured call	76–495 µs	28–220 µs (2.2–2.7×)
SSE, 64 KiB frame in 4 KiB writes	big tool result / long delta	139 µs	19 µs (7×)
SSE, small frames, 64-byte writes	normal streaming	63 µs	29 µs (2×)
Responses cleanup, 128 msgs w/ badge	Codex session history	2.3 ms	190 µs (12×)
Responses cleanup, nothing to strip	non-Weave history	51 µs	53 µs (parity)
Escalation, 2,048 fragments	long tool-heavy turn	3.3–12.2 ms	1.5–5.0 ms

Large inputs (the quadratic cases): tool-arg repair 36–134× faster at 16k fields; SSE 13× at 1 MiB; Responses cleanup 84–178× at 1k messages.

The one cost: tiny repairs (1–8 fields) pay ~1–3 µs of index setup. Everything else is neutral or faster.

Verification

• 15,766-case public-API corpus: zero differences from base (the earlier revision had 26).
• Full test suite, race suite, vet, lint; 3 fuzz targets × 60 s with seeds committed.
• 46-case SSE matrix: no case slower than base. Translation/escalation outliers re-checked with alternating run order; none reproduced.
• Weave checks: 18/19 pass; the diagnostic-logging findings are on pure JSON helpers with no external calls and were declined per that check's scope exclusion.

Local Docker was unavailable, so replay-smoke relies on CI.

🤖 Generated with Weave Router

[cursor]

Stale Bugbot comment from a previous run.

[greptile-apps]

The PR is not yet safe to merge because valid colon-prefixed schema properties can no longer receive deterministic argument repairs.

_{Reviews (1) · Last reviewed commit: "Remove quadratic router payload processi..."}

[cursor]

Cursor Bugbot has reviewed your changes using default effort and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 20325d6. Configure here.}