[schickling-assistant]

Summary

Closes #115.

Pin the main non-flake Python oracle to mergemaster v0.6.0, cff667c, and retire the temporary Fractal resync exception. Retain the retired Python parity-drift implementation and assertions through a separate immutable v0.5.0 oracle. Neither input enters the production package/runtime closure.

Head: 459cf58.
Base: cff667c.
Whole PR: 248 changed lines (76 additions, 172 deletions), eight files.
Delta after source GO at 0544b90: one new signed commit, 37 changed lines (32 additions, 5 deletions), four files. This delta requires review; the earlier GO does not cover it.

Main oracle and script guard

• python-oracle rev: cff667c; lastModified: 1794298270; narHash: sha256-UeHYs08Nzqy50cvlWu1n5Z0bR2I/J2Ectj7ztd5gfNI=.
• The original input-only lock operation changed only python-oracle; unrelated pins were unchanged.
• Remove all 29 resync-blob rows and the pending/resync header, together with their dual-source, expiry and retarget machinery. Every nonexempt script must match the new oracle's bytes and executable bits. Additions remain refused. Only scripts/PROVENANCE and scripts/test-scripts are exempt.
• Include inherited guard P3 deletion refusal: a reverse scan rejects oracle scripts missing from the working tree. Its footprint is six checker lines plus a four-line dedicated test (11 lines including separator), plus the README statement of bidirectional equality.
• Remove the oracle-scripts check's now-unused Fractal fixture environment dependency. Other fixture consumers are unchanged.
• Both v0.6.0 and the working tree lack scripts/parity-drift and scripts/parity-drift-test. Remove their obsolete runner restoration and the parity-only fixture-manifest provisioning. The runner discovers 25 suites from v0.6.0; every one exists locally. Guard tests refuse additions using either retired name.

Delta: retain parity-drift differential coverage

The first #120 CI run (42783562115) failed on both architectures because parity_drift.rs still loaded the retired Python suite from v0.6.0. Keep all its assertions and fixtures, rather than deleting coverage:

• Add non-flake python-oracle-parity at full SHA 1392a51; lastModified: 1793953851; narHash: sha256-NL/NWSn+YfpwP7h+CGZkeNmRh9hnQOH3OkF9GtyPXgM=.
• Expose MERGEMASTER_PYTHON_PARITY_ORACLE in checks.evidence and the development shell, the two environments that run Rust tests. Only crates/mergemaster-evidence/tests/parity_drift.rs switches variables. Its Python DRIVER and all assertions remain unchanged.
• The delta lock operation adds only python-oracle-parity and its root-input reference; all existing source pins remain unchanged.
• CI uses nix flake check and nix develop (.github/workflows/ci.yml:26-28), so it inherits the new variable without workflow-specific exports. This repository has no devenv configuration. The script-only guard does not consume the retired parity oracle.
• Evidence signer/verifier differentials keep MERGEMASTER_PYTHON_ORACLE at v0.6.0 (tests/gate.rs:357-360 and tests/generic_evidence.rs:232-235). The scheduled drift workflow continues to use the Rust CLI (.github/workflows/parity-drift.yml:33).

Concerns / deferred work

Janitor's provenance-table P3 remains for #118: scripts/PROVENANCE still has the e83871ec watermark and 29 stale table rows. Regenerating those rows and the watermark at Fractal e2721e52 costs at least 60 changed lines, which would take this 248-line PR above 280. No partial watermark/table rewrite was made. The strict script guard compares files directly to the pinned oracle and does not rely on this historical table; parity_drift.rs does read the watermark.

Evidence

Local plain Python, before the delta (unchanged guard code):

• python3 tests/test-oracle-scripts.py: 8 passed (0.010 s).
• Real working-tree guard against fetched v0.6.0 source: passed.
• JSON comparison confirmed the original lock update changed only python-oracle; subsequent comparison confirmed the delta adds only the parity input/root reference.
• Discovery confirmed the retired pair is absent in both v0.6.0/local trees and all 25 discovered suite scripts exist locally.

Local delta proof on dev3, outside 17:00-23:00Z:

• cargo fmt --all --check: passed.
• cargo test -p mergemaster-evidence --test parity_drift: passed (one Rust integration test; all 19 Python differential cases passed, 4.939 s). Ran under gate-slot --lane build --profile disk=cargo-warm; CARGO_TARGET_DIR=/var/tmp/mergemaster-oracle-repin-target, isolated TMPDIR beneath that target, GIT_CONFIG_GLOBAL=/dev/null and GIT_CONFIG_NOSYSTEM=1. Rust toolchain 1.98.1 and Python 3.14.7 came from the pinned shell's evaluated tool paths.
• The preceding cold run compiled successfully but failed one fixture because TMPDIR was inside this real Git checkout: the missing-default test found the enclosing repository's PROVENANCE. Moving only the fixture temporary directory outside the checkout fixed that environmental interference; no test assertion or source changed. An initial queued attempt was cancelled before admission to complete the linker PATH.
• Available /var space before the cold run: 114,139,070,464 bytes. Both Cargo invocations used the requested build lane and disk profiles. Gate-slot labels the run clean=no, so this is not claimed as an exact-head gate receipt; the tested source was signed head 459cf58 with no tracked worktree delta.

Input-only Nix operations used gate-slot --class eval --max-hold 600, one command per invocation, max-jobs=0 and builders='' (no builds). Lock addition and dev-shell attribute evaluation succeeded. print-dev-env was refused because realizing its shell environment would require a disabled local build; tool paths were obtained with nix eval instead.

NOT RUN for the new head: full Cargo suite, clippy, Nix builds/flake checks, full copied suite/genie, or hosted CI. No push was performed for the delta; prior hosted failures are not proof for the new head.

Posted on behalf of @schickling

field	value
agent_identity	agent/compoundingtech/mergemaster
session	01a19dbc
agent_persona	generalist
agent_supervisor	unavailable
agent_tool	OMP
agent_tool_version	18.12.5
agent_runtime	OMP 18.12.5
agent_model	anthropic/claude-opus-5-5
worktree	port-fractal-resync/port-oracle-repin-v0.6.0

[schickling-assistant]

SOURCE GO (ci-manager; independent GPT reviewer subagent, read-only through the GitHub API; no builds or tests run). 7 files (+49/-172). flake.lock updates only the python-oracle node, to cff667c (v0.6.0). The 29-row exception and the dual-source guard are removed completely. The blob IDs and modes of the copied scripts match the oracle at cff667c. Both retired parity tools are gone. The guard refuses additions, deletions, renames (also into exempt names) and symlink replacements. The workflows and the flake use the new guard and the Rust parity CLI and have no stale restore dependencies. Condition: hosted CI green and Janitor U04 before the merge.

q19-go v1 head=0544b90fab8b3d38cb5d14c0c271edc3235c6f8d verdict=GO

Posted on behalf of @schickling

field	value
agent_identity	agent/dotfiles/ci-manager
session	01a173fb
agent_persona	generalist
agent_supervisor	unavailable
agent_tool	OMP
agent_tool_version	18.12.5
agent_runtime	OMP 18.12.5
agent_model	anthropic/claude-opus-5-5
worktree	unknown

[schickling-assistant]

Delta GO (ci-manager; GPT reviewer subagent, cross-family to the Anthropic author) for 0544b90..459cf58. No findings. The delta is exactly 4 files, +32/-5: the separate non-flake oracle input pinned by full SHA 1392a51 (= v0.5.0); the env export in checks.evidence and in the dev shell; and parity_drift.rs, where only the variable and the error message change (DRIVER and assertions unchanged). The pin is justified: the v0.6.0 oracle retired scripts/parity-drift-test (the FileNotFoundError in run 42783562115). NOT RUN: local cargo, clippy, Nix checks. Hosted CI 42794311341 was in progress at review time; merge needs it green.

q19-go v1 head=459cf58193e0bf1d0c9c69222beb5f011f01d730 verdict=GO

Posted on behalf of @schickling

field	value
agent_identity	agent/dotfiles/ci-manager
session	01a173fb
agent_persona	generalist
agent_supervisor	unavailable
agent_tool	OMP
agent_tool_version	18.12.5
agent_runtime	OMP 18.12.5
agent_model	anthropic/claude-opus-5-5
worktree	unknown

[schickling-assistant]

U04 exposure audit: PASS at exact head 459cf58193e0bf1d0c9c69222beb5f011f01d730 (oracle re-pin to v0.6.0 plus a separate parity oracle). This replaces the earlier HELD on red CI at 0544b90f.

• CI 42794311341 succeeded (pull_request, PR list contains Pin Python oracle to v0.6.0 and retire resync exceptions #120). All four jobs succeeded: checks and publisher-stack on ubuntu-24.04 and ubuntu-24.04-arm. Release 42794311422 succeeded.
• Hosted test evidence: both checks jobs printed test python_oracle_parity ... ok. This is the test that failed at 0544b90f.
• Both commits are verified. The LICENSE equals main. A merge-tree against main cff667cd53bdb6c024f9d7a06259b8f9a3fbb0cb is clean.
• Delta review (0544b90f..459cf581, 4 files):
  • The new non-flake input python-oracle-parity pins compoundingtech/mergemaster at 1392a516e969be46d9516d904b76a4db98c55030 with a locked narHash. That commit is on main and is contained in tag v0.5.0.
  • The input is exposed only as MERGEMASTER_PYTHON_PARITY_ORACLE in checks.evidence and the dev shell. Only crates/mergemaster-evidence/tests/parity_drift.rs reads it; the test driver and assertions are unchanged. No runtime package or binary depends on it.
  • python-oracle stays at v0.6.0, so the bidirectional oracle check of this PR is kept.
  • No workflow, permission or new third-party dependency change.
• P3, carried from the earlier review and now tracked in Validate tests/fixtures/fractal PROVENANCE blob IDs against the pinned fixture #118: scripts/PROVENANCE still names commit e83871ec, and 29 table rows are stale. This has no exposure impact.

Signed: Janitor (agent/dotfiles/janitor).

Posted on behalf of @schickling

field	value
agent_identity	agent/dotfiles/janitor
session	01a1a1ea
agent_persona	generalist
agent_supervisor	unavailable
agent_tool	OMP
agent_tool_version	18.12.5
agent_runtime	OMP 18.12.5
agent_model	anthropic/claude-opus-5-5
worktree	unknown