Add bounded Camel/KIE CPU fabric and cached Chrome corpus summarization

[hsoliwal]

Outcome

Adds a bounded CPU execution fabric, a fixed-loopback Chrome2api inference lane, Chrome Summarizer-compatible single/document modes, and content-deduplicated corpus summarization to CodexPro's existing fabric MCP tool.

Runtime surfaces

• invariants: recursive fleet capability laws
• dag_contract / dag_execute: Drools-admitted, Camel-executed data-only CPU DAGs
• chrome_contract / chrome_status / chrome_complete: bounded loopback ChromeML text inference with deterministic receipts
• chrome_summarizer_contract / chrome_summarize: normalized Chrome summarizer option surface
• chrome_summarize_document: exact UTF-8 chunk/map/reduce document summarization
• chrome_summarize_corpus: safe top-level workspace glob, exact-content deduplication, per-path alias receipts, and sequential unique-content execution

Corpus and cache boundary

• at most 64 files and 8,000,000 total UTF-8 bytes; each source remains subject to the 2 MB document cap
• rejects path separators and recursive ** globs; reuses the existing workspace path guard and file inventory
• sorts document IDs and content hashes for deterministic execution and failure order
• duplicate bytes share one content summary while retaining distinct path-bound alias receipts
• successful summaries use a 128-entry / 8 MB process-local LRU cache keyed by endpoint, exact source root, and normalized options root
• concurrent identical work joins one provider request; failures are never cached
• cache diagnostics do not alter semantic receipt identity
• summaries remain authority_ceiling=CANDIDATE_ONLY

Chrome boundary

• wraps xszwow/Chrome2api commit 406c5c7a33a6c364c0bd42892250ad1b6a103dad
• fixed literal-loopback /v1 endpoint; callers cannot provide a URL
• no forwarded authorization, cookies, browser profile, escaped file path, media, redirect, or SSE surface
• no Chrome runner, proprietary DLL, or model weight is copied or redistributed
• compatibility backend is truthfully marked non-native and batch-only

Camel/KIE boundary

• Apache Camel 4.22.1 LTS with Drools 10.1.0 fail-closed admission
• four allowlisted pure capabilities
• bounded SEDA queues/workers, stable fan-in, and replayable receipts
• model plans are data-only and cannot provide endpoint URIs, shell, executables, or credentials

Proof on 090aa1abcdc08c9625b652131d532922f021b526

• full npm run release:check: passed at codexpro@0.30.2
• Maven: 6 tests passed
• complete TypeScript/MCP/CLI/HTTP/widget/release smoke suite: passed
• Chrome Summarizer option matrix: 72 combinations passed
• corpus discriminator: 3 paths / 2 unique contents / 2 initial provider calls / 0 replay calls
• source-blind stdio MCP behavior contract: 4/4 clauses passed, including restart cache reset and adversarial glob/empty-source checks
• stress suite: passed
• npm audit --audit-level=high: 0 vulnerabilities
• dry-run pack and clean tarball install: passed; all three public CLIs report version 0.30.2
• real E:\Downloads\*AI*.md inventory: 14 files / 7 unique contents / 2,085,697 bytes / corpus root e4927f03b3cc65855f6d7367e8b37ab58228590d0848f1d83b8a5d62391e84ab

The branch also merges current main and hardens its package-install smoke with an explicit private consumer manifest, preventing npm from escaping to an ancestor project.

FOSS_REUSE_DECISION.tsv and THIRD_PARTY_NOTICES.md pin the evaluated donors, decisions, licenses, leaf hashes, and proof status.

Runtime dependency

Real summaries still require the separately installed Chrome2api loopback service and Chrome-managed model assets. This repository does not redistribute those proprietary assets.