Redact capture fields before serialization instead of post-hoc string surgery

[efenocchi]

Follow-up from the independent review of #360 (see its final review comment).

Every capturer runs redactSecrets(JSON.stringify(entry)) (src/hooks/capture.ts:177, the codex/cursor/hermes hooks, and src/mcp/cowork-ingest.ts since #359). tool_input / tool_response are themselves JSON.stringify-ed first (src/hooks/capture.ts:134-135), so the redactor sees doubly-nested JSON text. #360 keeps the common single-level trailing-escape valid (maskBeforeQuote in src/hooks/shared/redact.ts), but a secret that is purely backslashes or ends in a literal " under that double nesting still produces an unparseable line, which the session queue stores as an opaque raw_message row. Pre-existing on main; identical before #360.

Redacting serialized text cannot be nesting-correct in general (the inner escape \" is \\\" at depth 2, 7 chars at depth 3). The sound fix is to redact the string fields (content, tool_input, tool_response) before serialization in each capturer, then serialize once. Regression shapes to keep: password=\\, password=abc", token=xy\z\, nested through JSON.stringify({ content: JSON.stringify(secret) }).

Also from CodeRabbit on #360 (same area, separate behaviour change): quote-aware tokenization so password="two words" masks the whole quoted value; today the rule stops at whitespace.

[sumitvairagar]

Opened PR #363 with the structural fix.

The approach: redact content, tool_input, and tool_response individually at one level of JSON encoding before building the entry, then JSON.stringify the entry once — no post-hoc string surgery over doubly-encoded values. All four regression shapes from the issue (password=\\\\, password=abc\", token=xy\\z\\, nested through JSON.stringify) are covered by new tests. Also includes the quoted multi-word rule (password="two words") from the CodeRabbit note on #360.