Repository navigation
oracle: remove the historical Mall item files from the public tree (#100) - #101
Conversation
) Removes oracle/corpus-items (859 files, 69,552,646 bytes) and oracle/cm5-cmd-ops-baseline.json (8,666,084 bytes; it embeds about 4.2 MB of corpus-items text). They have no redistribution permission (Ascendance3D/X_ITE_R#168). Both are rows of the new oracle/private-inputs.tsv and are gitignored. cm-evidence gains private_input: $CM_PRIVATE_INPUTS/<name>, else oracle/<name>, else the main worktree's; each is verified by kind, file count, bytes and sha256 (the X_ITE_R answer-key measure). Absent: fails in strict mode; skips in ci mode only when CM_EVIDENCE_ALLOW names the input, and logs the skip. Mismatched: always fails. The cm-modes baseline check reads corpus files through it (the cases and the other files still run when the corpus is skipped); cmd_fuzz_oracle reads its baseline and start documents through it. CI allows the two inputs by name, and tools/evidence_summary.py checks the keyed reason. No assertion changes; no baseline is re-recorded. A verified private backup is outside the repository. Git history is not rewritten.
An input path whose symlink target does not exist was treated as absent, so ci mode skipped it. It now fails as a mismatch, as any unreadable entry does. Unit tests cover the broken symlink, an extra file and a file where a directory is wanted.
Independent review and owner decision (2026-10-29)MiniMax: failed on its context limit. This is not a passed review. Independent reviewer: a Claude subagent with a fresh context. It reviewed the 13 non-corpus files at
Confirmed correct by running: strict mode fails on each absent input; a mismatch (extra file, wrong kind, truncated file, one changed byte) fails in every mode; the allow list matches names exactly; every permitted skip writes one log line; the original assertions are unchanged; the corpus digest reproduces with Public CI coverage lost (logged skips, not passes): Owner decision: the owner approved Mandatory Reverse Engineer Anything (REA) Evidence — Not Applicable: asset removal and test input resolution only. No JavaScript-to-Rust behavior changed. The CodeMirror 5.65.21 pin in |
Summary
This PR removes the historical Mall item files from the public tree. They have no redistribution permission. They are the same files as in Ascendance3D/X_ITE_R#168.
oracle/corpus-items/0a8127f2…b66f534(tree listing)oracle/cm5-cmd-ops-baseline.json(embeds about 4.2 MB of corpus text)59a1d53e…27ed5Primary issue: Closes #100.
oracle/private-inputs.tsv(new) lists both paths, with kind, file count, bytes and sha256. Both paths are gitignored.cm-evidence::private_input(new) finds and verifies each input. It looks in$CM_PRIVATE_INPUTS/<name>when the variable is set. Otherwise it looks inoracle/<name>, then in the main worktree. A missing input fails in strict mode. In CI mode it skips only whenCM_EVIDENCE_ALLOWnames the input itself, and each skip is logged. A mismatched input always fails.cm-modesbaseline check reads corpus files through the resolver. When the corpus is skipped, the cases and the non-corpus files still run.cmd_fuzz_oraclereads its baseline and its start documents through the resolver.tools/evidence_summary.pyaccepts the keyed reasonprivate-input-absentonly for an allowed key. It always refusesprivate-input-mismatch.Kept after a check: the
oracle/corpus-items/...file rows incm5-vrml-baseline.jsonandcm5-javascript-baseline.jsonhold a path, counts and sha256 digests only.docs/provenance/corpus-items-ffe080b.sha256holds hashes and paths only.cm5-doc-ops-baseline.jsonandoracle/cm5-modes-inputs/are synthetic.Owner decision: new CI allow-list entries
CI has no copy of the private inputs. This PR adds two names to
CM_EVIDENCE_ALLOWin.github/workflows/rust-tests.yml:corpus-itemsandcm5-cmd-ops-baseline.json. The CI run then skips by name, and logs:vrml_tokens_match_js_vrml_mode: the 606 corpus files. Its 531 cases still run.javascript_tokens_match_cm5_runmode: the 253 corpus files. Its 821 cases and 294 upstream files still run.cmd_fuzz_oracle: the whole test, because its baseline is private.The strict local suite still runs all three in full. Please approve or reject these entries. This PR is not merged until you decide.
Verification
tools/fetch-upstream.sh, thencargo test --workspace --no-fail-fast,CM_EVIDENCEunset): 803 passed, 0 failed, 0 ignored. This includes 3 newcm-evidencetests.vrml531 cases + 606 files, 0 mismatches.javascript821 cases + 547 files, 0 mismatches.cmd fuzz600 sequences, 22,189 steps, 0 mismatching (2 known). The samevrmlresult comes fromCM_PRIVATE_INPUTSpointed at the private backup.evidence private-input-absent (corpus-items)or(cm5-cmd-ops-baseline.json). With the old CI allow list (no-gpu-adapter,tool-absent), they also fail.vrmlruns 531 cases + 0 files, andjavascriptruns 821 cases + 294 files.tools/evidence_summary.pyexits 0.cargo fmt --all --check: clean.Private evidence
A verified private backup is outside the repository: 860 files, 78,218,730 bytes, and every file is in its
SHA256SUMS. It equals the Git blobs at32f3cca, and it is byte-identical to the X_ITE_R #168 backup of the same files. It is not committed.Limits
Removal from
masterdoes not erase the files from Git history. Older commits, forks, clones and GitHub caches still hold them. Erasure needs a history rewrite and separate owner authorization.CLAUDE.mdsays "Keep the repository PRIVATE", but the repository is public. This PR does not change visibility or that rule.Mandatory Reverse Engineer Anything (REA) Evidence — Not Applicable
This PR removes assets and changes how tests find their inputs. No CodeMirror behavior changes. The CodeMirror 5.65.21 pin, every baseline and every assertion are unchanged.