Add structured request-content parsing to Egress Gate - #35

Summary

• add reusable request-content parsers for complete UTF-8 bodies, selected JSON fields, and normalized message blocks
• let the regex gate detect, deny, and source-preservingly replace selected structured content while retaining complete-body behavior
• enforce the 4 MiB input/output boundary directly at the public parser API
• preserve distinct message-block classifications through filtering while deduplicating identical (node, kind) classifications
• document the parser architecture, configuration, limits, overlap semantics, and replacement behavior

Why

Egress Gate previously treated a request body as one complete UTF-8 string. Agent harness requests need bounded, configurable extraction of model-visible JSON fields and message blocks without scanning unrelated metadata or losing exact source formatting during replacement.

QA also found two boundary issues in the first implementation: the reusable parser API relied on downstream request models for some body-size enforcement, and a broad text selector could shadow an explicit tool classification before block-kind filters ran. The final implementation owns its advertised bounds and filters distinct classifications before exposing each backing string once.

Impact

Policies can opt into json-fields or message-blocks scans. Existing body, path, query, and header behavior remains available. Structured replacement rewrites only selected JSON string tokens and preserves all unrelated request bytes. Explicit tool-input and tool-output filters cannot be bypassed by an overlapping generic text selector.

QA and validation

• make check — 324 tests plus formatting, linting, typing, installed-package import, and dependency audit
• Python 3.11 focused parser/regex suite — 83 tests
• python3 tests/test_render_dev_notes.py — 11 tests
• scripts/build-docs.sh — strict build, 14 agent-readable pages, all documentation tests pass
• isolated adversarial parser QA — 2,000 generated valid documents, 25,000 strict-JSON mutation comparisons, 4,000 concurrent parses, exact resource boundaries, replacement ordering, and classification-limit probes
• isolated end-to-end QA — source and installed-wheel CLI, generated schema, policy validation/evaluation, direct processor, live gRPC error/recovery and concurrency, all scan/action variants, and pipeline chaining
• isolated documentation/adoption QA — documented snippets and policies, installed workflows, strict site build, served HTML/Markdown routes, schema/runtime cross-checks, and limit/error wording
• final multi-agent rerun — parser, end-to-end, and documentation lenses all clean after fixes