perf(parser): single-pass field extraction in the large-line JSONL scanner

[iamtoruk]

Follow-up to the desktop-slowness work: the large-line JSONL scanner was ~55% of warm-run CPU because parseLargeJsonl extracted each top-level field with an independent findObjectFieldValue walk — a multi-KB tool blob preceding the target keys got re-scanned up to ~8-9 times per assistant line.

• New extractObjectFields: one walk over the object captures all requested fields (capture-or-skip, early exit when all found). Applied at the root (6 fields) and assistant-message (4 fields) levels. Exact semantics of the per-field walks preserved: raw-key comparison, first-match-wins on duplicate keys, identical behavior at truncated-key/unparseable-value failure points, same resync on stray content.
• The old find* helpers remain — still used by usage/tool-block/user-text extraction paths.
• Cache-load scoping was investigated and deliberately NOT done: every provider's incremental section (including durable OTel orphans) lives in the one session-cache file, so skipping the load on filtered runs would force cold re-parses and drop orphan history.

Evidence:

• Synthetic 420MB corpus (large blobs preceding keys): cold parse 48.6s -> 23.3s (~2x).
• Real 21MB session file, cold parse: 42.1s -> 17.4s (2.4x), output byte-identical (verified independently).
• CPU profile: the per-field amplification frames drop to ~0%; remaining cost is the inherent single container scans.
• 13 new tests (12 characterization pinned against the OLD implementation first, 1 whole-pipeline aggregate invariant); parser suites green.