feat: installer error-report generation (schema v1)

[forkni]

Summary

• Add sd_installer/report.py: builds a self-contained diagnostic report (schema v1) on installation phase failure, with an env-var allowlist/denylist so secrets never leak into the report
• sd_installer/installer.py: track current_phase/_last_pip_stderr, add _write_install_error_report(), wrap the phase-calling loop in try/finally so a failure always writes a report before re-raising
• sd_installer/cli.py: add a report subcommand

This is PR I-2, chained on deps/cuda-link-1.12-torch-modelopt-onnx (PR I-1, #1) because _write_install_error_report() reads current_phase values set by phase4b_cuda_link/phase4c_cuda_link_env, which only exist on that branch, not on main. Once PR I-1 merges, this PR's base should be retargeted to main.

Test plan

• python -m sd_installer report writes an install_error_report_*.txt
• A forced phase failure writes a report and re-raises via the new try/finally path
• Claude Code review (requires CLAUDE_CODE_OAUTH_TOKEN secret on this fork)
• GitHub Copilot review

[claude]

Claude finished @forkni's task in 5m 15s —— View job

---

Code review — PR #2 (feat: installer error-report generation (schema v1))

Todo list

• Gather context (PR diff, CLAUDE.md files)
• Check precondition (not closed/draft/trivial/already reviewed)
• Summarize PR changes
• Run parallel review (CLAUDE.md compliance x2, bug scan x2)
• Validate flagged issues
• Post final review summary

Summary of changes

• sd_installer/report.py (new, 184 lines): builds a schema-v1 diagnostic report with an env-var allowlist (CUDALINK_, HF_, SD_, SDTD_ prefixes) filtered further by a denylist (TOKEN, KEY, SECRET, PASSWORD, PASSWD, CRED), so HF_TOKEN/SD_API_KEY-style vars are dropped despite matching an allowed prefix.
• sd_installer/installer.py: adds current_phase/_last_pip_stderr tracking, wraps the phase-calling loop in try/except so any phase failure writes a best-effort report via _write_install_error_report() before re-raising the original exception.
• sd_installer/cli.py: adds a report subcommand for generating a report on demand (no failure required).

No CLAUDE.md files exist in this repository, so there was nothing to check for compliance.

Bug/security review

Two independent reviews focused on the diff itself (correctness, secret leakage into reports/logs, and whether error-reporting stays best-effort) came back clean. Specifically verified:

• args.base_folder used in the new cmd_report (cli.py) resolves correctly — --base-folder is a shared top-level parser argument (cli.py:396), inherited by every subcommand, matching the existing pattern in all other cmd_* handlers.
• The env allowlist/denylist in report.py correctly drops secret-shaped vars (e.g. HF_TOKEN, SD_AWS_SECRET_ACCESS_KEY) even though they match an allowed prefix — full os.environ is never dumped.
• _write_install_error_report and write_error_report are both wrapped in broad try/except that print-and-swallow, so a bug in reporting can't mask the real install failure; the phase loop still re-raises the original exception in all cases.
• The phase-loop refactor in installer.install() preserves the original phase order/semantics; success is only read on the non-exception path, so there's no unbound-variable risk on the raise path.
• phase4c_cuda_link_env's new SDTD_BASE_FOLDER_PATH persistence is Windows-gated, independent of the other setx calls, and non-fatal on failure — consistent with existing CUDALINK_* handling.

One thing noted but not flagged as a defect: the TRACEBACK/LOG TAIL sections dump raw exception text and pip stderr verbatim (unlike the ENV section, these aren't allowlist-filtered). In this repo all pip index URLs are public (PyTorch/NVIDIA/NGC), so there's no credential-bearing URL in the current pip invocations — this isn't an active leak path today, just worth keeping in mind if a private/authenticated index URL is ever added later.

No issues found. Checked for bugs, security (secret leakage), and best-effort error-handling contract compliance.

[forkni]

Re-triggering review with bumped max-turns (60 -> 120) — prior run exhausted its turn budget mid-review.