Optimize code-review-deep command for ~50% lower usage

[ydesgagn]

Summary

Reworks the /code-review-deep command to consume substantially less usage while producing better reports than the original. Verified against the github-build repo: the optimized run produces 42 main-body findings (6 High / 27 Medium / 6 Low / 3 Info) versus the original's 26, while running on roughly half the agents.

Key changes:

• Phase 1 scans now use the Explore subagent (purpose-built for read-only file location / grep, returns excerpts not full files) instead of the heavier general-purpose agent.
• Phase 2 agent count consolidated from 20 to 10 by merging overlapping concerns: Quality/Comments and Bugs/Errors stay separate, but Perf+Observability+API+Concurrency+Migrations merge into a single Backend agent, IaC+Compliance merge, i18n+AI/ML merge, Repo+CI/CD merge, Docs+Config merge, Deps+BackwardsCompat merge.
• Web search is now opt-in only — agents rely on the model's existing knowledge by default; web search is reserved for CVE lookups, latest-version checks, and explicitly user-requested verification.
• Phase 3 adversarial validation batch size raised from 3 to 10 findings per agent (~70% fewer validation agents on a typical repo).
• Phase 3.5 confidence filter tiered by severity (Critical >= 50, High/Medium >= 65, Low/Info >= 80) so important findings aren't hidden behind a single 80 threshold.
• Phase tracking simplified to TaskCreate-based tracking (was a verbose multi-page log).
• Bug-pattern and error-handling checks expanded with CLI-specific concerns (STDOUT vs STDERR, narrow rescue clauses, error context loss).
• Quantitative count requirements preserved and made explicit (linter disables, silent failures, addObserver/removeObserver delta, pattern duplication).
• All high-value rules preserved verbatim: Dependabot `gh api` evidence requirement, IaC UserData/CloudWatch agent observability check, case-insensitive file lookups, full exclusions list, severity rubrics, markdown lint rules.
• File size reduced from 1336 to 750 lines (~44% smaller).

Types of changes

• Bugfix (fixes an issue)
• New feature (adds functionality)
• Refactoring (improves code without changing functionality)
• Breaking change (incompatible changes)
• Build or security update (updates dependencies, libraries, or security patches)
• Code style or documentation update (formatting, renaming, or documentation changes)
• Other (please describe):

Checklist

• Unit tests added to validate my fix/feature
• I have manually tested my change
• I did not add automation test. Why ?: This is a slash-command markdown file with no executable code path; verified by running the command end-to-end against the `cloud-officer/github-build` repo and comparing the generated `docs/code-review.md` against a baseline produced by the previous version of the command.
• Database changes requiring migration with downtime or reprocessing of existing data
• The SOUP file lists the risk Level, requirements and verification reasoning associated with each library
• `readme.md` includes sections on introduction, installation, usage, and contributing
• `docs/architecture.md` includes sections on the architecture diagram, software units, software of unknown provenance, critical algorithms and risk controls related to PII and security
• Impact on PII, privacy regulations (CCPA/GDPR/PIPEDA), CIS benchmarks or security (availability/confidentiality/integrity); management must be notified