fix: bound GitHub pagination and credential forwarding

[YuzeJ21]

Summary

• validate every GitHub pagination target before use and reissue accepted targets only to the verified HTTPS api.github.com origin
• bind file and commit traversal to the requested repository, pull request, collection, and GitHub's numeric repository alias only when that alias matches verified PR metadata
• enforce deterministic per-fetch request, page, item, decoded-response, file, and commit ceilings
• stop file and commit retrieval at the smallest observable overflow while preserving API order and marking incomplete evidence partial
• reject ambiguous, malformed, escaped, cyclic, repeated, or over-budget pagination fail closed
• preserve saved reviews, reports, exports, and Streamlit session state when pagination ingestion fails

Review-driven correction

The first independent review found that real GitHub pagination may canonicalize named repository paths to /repositories/{id}/.... A live public API response reproduced that behavior. Focused failing regressions were added for file and commit pages plus a mismatched numeric repository ID, then the validator was corrected to accept only the numeric ID bound to verified pull-request metadata. The complete verification below was rerun on the repaired head.

Verification

• Ruff: passed
• complete suite: 2,049 passed, 2 intentional skips
• combined coverage: 95.24%
• repository contracts: 84 passed
• deterministic acceptance benchmark: 12/12 cases, zero mismatches, zero must-have False Ready outcomes, zero false blockers
• comparison benchmark: 2/2 cases, zero mismatches; Stage 1 unchanged
• two byte-identical wheel builds: SHA-256 07633f74035559e085cdf18839a4b966cf7bbfffa466ebed15b9f1b15c496b01
• artifact inventory: 101 entries, zero forbidden entries
• clean installed-wheel dependency, version, CLI, benchmark, and loopback workbench-health checks: passed
• installed-wheel packaged Chromium regression: 3 passed
• local CPython 3.11.15 and 3.13.14 lanes: complete suites passed before the final canonical-path repair; the repair's focused tests, exact repaired wheel installs, dependency checks, versions, and installed benchmarks passed in both lanes
• independent review: the initial Important finding was corrected test-first; exact repaired-head re-review passed its focused tests and lint with no remaining reported finding

Evidence boundary

This is engineering evidence only. ScopeProof remains an evidence assistant, not a correctness oracle; it never executes target-repository code. Partial ingestion remains fail closed, persisted/exported objects remain Pydantic-validated, and the opt-in GitHub Action remains informational. No release, tag, package publication, outreach, participant contact, benchmark retuning, or product-stage advancement is included.

Product Stage 1 remains exactly:

• 0/5 qualifying reviews
• 0/3 independent practitioners
• 0/3 public repositories
• 0/3 independently observed under-ten-minute completions
• 0/2 reuse-intent signals

Unsupported environments remain real screen-reader operation, Windows desktop workflow, Linux desktop workflow, non-Chromium browser behavior, and accessibility conformance.