Wire governed SI execution, bounded verification, and repair continuation

[infotradescout]

SI's controls must govern real execution so a task cannot be declared complete after empty tests, a no-op write, a detached repair, or a substituted acceptance command.

This draft connects the existing controls to the runtime and completes the repair continuation:

• Persist checkpoint-bound retrieval windows across restarts and concurrent callers, enforce the 12-file default and 65,536-byte complete-packet ceiling, and retain excluded dependency identities in coverage.
• Check prospective source changes through the project index, reject new exact duplicate implementations and named ownership conflicts, and preserve inherited debt and approved generated projections.
• Observe unchanged files without inventing write receipts or resetting the handoff budget.
• Bind repair verification to the exact original failing command; carry authentic bounded failure output to the repair worker; propagate verified nested repair completion to its original task.
• Support a bounded native Node test command alongside Python verification. Reject unsupported flags, paths, symlinks, and zero-test false success.
• Regenerate the existing ChatGPT adapter and retain the canonical product and responsibility documentation.

Validation:

• TMPDIR=<writable-external-temp> PYTHONDONTWRITEBYTECODE=1 python3 -B skills/selective-intelligence/scripts/quality_gate.py --output <receipt.json>: all five required gates passed, including 334 unit tests.
• Discovery bridge: 11 passed; native pointers: 7 passed; ChatGPT adapter: passed (59 files); public plugin: 4 passed.
• Independent Objector review passed on local candidate 0b312f58e0faab6b633106918ab305ab491f377f, including repair-lineage and command-preservation checks.
• Published commit: a810deebca2a36decf1bb6238aa5cbc7f2d9f269. Its tree 96c54766d4fe5b304dc57010fda52a25693b01b0 was read back from GitHub and is identical to the reviewed local tree. The connected GitHub app supplied publication commit metadata because local Git HTTPS has no credentials.
• The local quality receipt records a clean tracked source, 320 source files, and source snapshot SHA-256 b15563a2a466524f893b639140c08f7da6581ac1afb9c8dfb56b42a766a6f4cc.

Limits:
This proves deterministic source/runtime controls when callers use the engine. It does not prove installation in ChatGPT, Codex, Windows, or Platynum; actual model dispatch; unrelated tool governance; live provider behavior; or full-stack delivery in a real client. Model behavior evidence remains not_run. No main merge or production release is included.

Current hosted Quality run 34774579873 failed before any test steps were recorded. Job metadata has runner_id 0 and an empty runner name; the cause is unconfirmed. The passing results above are local evidence, not hosted CI success. This PR remains a draft.

Independent hosted proof (2026-09-13): verified package summary passed on Render deploy dep-dajf8njm8hqs7382oe6g, using Python 3.12.14 and the checksum-pinned official Node v22.23.2 binary. The committed launcher a1db7ac6a0be27a6024cdb37cd9e4d3a7314ab08 recorded all candidate commit/tree identities and passed all 28 execution steps. Public summary identity was checked after deployment; source, archive, Git configuration, and build-script paths returned 404. This independent hosted proof does not clear or replace canonical GitHub required checks. The exact SI candidate passed all five canonical quality gates plus discovery, native-pointer, ChatGPT-adapter, and public-plugin checks.

[infotradescout]

Actual Windows runtime repair published — 2026-09-16

Current head aa787adfb696e493e2656fa853041eb294599ec8, normal fast-forward from a810deebca2a36decf1bb6238aa5cbc7f2d9f269; remote head read back after push. No main merge or global skill replacement.

On the now-accessible TSCommandCenter, the existing C:\Users\flavo.agents\skills\selective-intelligence install reports1.0.0 and differs from this1.0.8 runtime. Existing user work and that installation were preserved. Initial exact-candidate Windows quality failed unit_tests and release_integrity; three other canonical gates passed. The explicit baseline unit run recorded334 tests,9 failures,22 errors,1 skip.

Causal product fix

The real guarded writer used Path.write_text with platform newline conversion, while its write receipt and transactional staging check expected the supplied UTF8 bytes. On Windows, valid multiline artifacts were rejected as 'staged bytes differ from the supplied file content'. The canonical writer now uses newline="". No receipt, ownership, authority, staging or verification assertion was removed.

Added an actual apply/persist/receipt regression with mixed LF/CRLF/CR and Unicode in the existing transaction suite. Original writer: fails with the demonstrated staging error. Repaired writer: passes and verifies exact on-disk bytes plus the saved artifact SHA256. The existing mode-preservation test now compares supported before/after mode bits while retaining the literal0640 assertion on POSIX. Regenerated the59-file ChatGPT adapter and metadata from the canonical source; this is package generation, not ChatGPT installation.

Exact clean-source acceptance

Committed four files,21 additions/5 deletions. Source snapshot 7d82f184bc111c825762ed9fcb715f2fb01b9fba1e389af135b77984b54080cd,320 files, workingTreeDirty false, exact aa787ad head.

On the user's WSL Ubuntu environment (Python3.12.3; checksum-verified official Node22.23.2): all five canonical quality gates passed, including the current335-test unit suite; tools/test_chatgpt_adapter.py passed for59 files. Receipt observedAt 2026-09-16T19:05:28.460210Z.

Command used native external Linux temp directories for both TMPDIR and SI_EVAL_TEMP_PARENT. The first WSL attempt let the eval helper use its default Windows-mounted directory and correctly failed a POSIX directory-permission assertion. Only the test temp location was corrected; those control assertions were not relaxed.

Remaining limits — not global Windows readiness

The full Windows post-repair run still records335 tests with7 failures,5 errors and1 skip. Remaining failures include unavailable native symlink privileges, other test-fixture byte behavior and CLI/reuse cases; they were NOT hidden or relabeled passing. No symlink privilege, host security setting, global install, main branch or production application changed. Full Windows/client acceptance and installed cross-session operation remain separate work.

Local evidence root: D:\si-seo-recovery-20260916\evidence; authoritative files si-final-linux-quality.json/.log, si-newline-baseline.log, si-newline-candidate.log, si-windows-after-newline.log, si-adapter-regeneration.log. Next: bounded installed-runtime save/resume/action proof and remaining Windows-specific acceptance, not another repository-wide audit or repeat of this source repair.

[infotradescout]

Scoped installation and fresh-process resume now verified — September 16, 2026

The exact published aa787adfb696e493e2656fa853041eb294599ec8 skill was installed into a versioned, user-owned WSL location on TSCommandCenter: /home/flavorgood/.local/share/selective-intelligence/releases/aa787adfb696e493e2656fa853041eb294599ec8/selective-intelligence.

All117 installed source files were byte-verified against the accepted source. Installation manifest SHA256: 71ac4f4ad2248d078bf9fb29c2b4d172c03073d3bfe6edc9dcb92992bff4335b. No existing global Windows1.0.0 skill directory, default client pointer, security privilege or host configuration was replaced.

Executed the INSTALLED build_engine.py start command in a separate acceptance workspace. It saved session si-6197351c45264174a75485af9e7fa264, checkpoint cp-7248a228d914, intent hash 400118f0635e878666cc935e7ef1639cd695fef31c4a86810e77001ca62b1e90. A NEW OS process then ran the installed engine's show command using that persisted session. Session/objective/checkpoint/intent identities and mutation/execution locks matched exactly. No repository rediscovery was performed by the read-only resume.

Scope matters: This proves the versioned WSL installation runs and persists/resumes its locked checkpoint. No approved worker task, actual model dispatch, cross-project resume→action cycle or automatic ChatGPT/Codex/global Windows adoption was executed. It does NOT clear the remaining native Windows failures or establish Infinity installation. The checkpoint stayed proposed and executionLocked true; no approval was fabricated.

Evidence: D:\si-seo-recovery-20260916\evidence\si-scoped-install-start.json and si-scoped-install-resume.json. Continue from this installed revision and saved session; do not reinstall it, reset the global skill, or repeat source quality tests absent a relevant change.