Audit whether the canonical branch preserves all validated OpenSolar work

[Coconut-ch1ken]

Purpose

This draft PR exposes the current local canonical openJiuwen-Solar state for an independent preservation audit. It is not a claim that every Phase 22 limitation is fixed, and it should not be merged until the audit is complete.

Exact state under review

• Head commit: 81349555c85ce6b5a40899af79adf2dacdc955f1
• Base branch: openJiuwen-Solar
• The base tip was freshly fetched and verified as an ancestor of the head.
• The head contains the audited overwritten-contribution recovery and the three accepted mappings recorded in docs/integrations/autosci/phase-23-integration-ledger.md.

Review requested

Please use an independent agent to determine:

1. Whether useful committed work from the relevant local and remote branches is present in the head tree.
2. Whether obsolete files or duplicate pre-refactor paths were accidentally restored, especially harness/tests/....
3. Whether moved, semantic, and exact preservation classifications are supported by the current files and Git history.
4. Whether the Stellven README contribution and other trust-sensitive contributions remain intact.
5. Whether any ledger entry marked superseded, obsolete, or rejected still contains a useful change missing from this PR.

Important boundaries

• This PR intentionally excludes all untracked local reports, raw evidence, spreadsheets, and scratch files.
• Uncommitted changes in other worktrees are not represented here. In particular, .severity-repair-worktrees/p0-research contains uncommitted Windows/UTF-8/timeout/long-path edits that still require separate adjudication.
• The Phase 22 evidence paths under outputs/ and .codex-tmp/ are generally local and ignored; do not infer reproducibility merely from report references.
• P22-044, P22-045, P22-047, P22-054, P22-069, and P22-071 still require the Phase 23 semantic/evidence audit described in the current logs.

Validation available

• Required ancestry checks passed for both fetched remote openJiuwen-Solar tips.
• Accepted integrated hashes 55ec2ea78, 7d7ae2a1d, and bc3e9ff59 are reachable from the head.
• Focused current-session verification passed: 18 deterministic issue tests and 8 revision/delivery tests.
• An exact revise_required smoke completed with revision applied and final acceptance true.

Please report findings first, with commit/file references, and distinguish missing committed work from intentionally excluded uncommitted or ignored evidence.

[suraj-subrahmanyan]

Checked against openJiuwen-Solar by hashing every blob in both trees and tracing all 519 deletions
directly, independently of the ledger.

245 deleted files still exist byte-identical elsewhere, 244
exist under the same name with edits, and 29 of the remaining 30 are the hyphen-to-underscore
renames into tests/. All 29 targets were confirmed present with matching line counts. The last is
a regenerated CSS asset. Nothing committed is lost.

On the harness/tests/... question: the duplicate tree is removed correctly, and rescuing the 39
orphans by rename rather than deletion was the right call.

Three things worth fixing before merge:

1. harness/tests/ ends at 9 files, not 0. Nine new files are added back into the directory the
PR is retiring (test_self_rag.py, test_curriculum_loop.py, test_routing_bandit.py, and six
more). pytest.ini sets testpaths = tests, so none of them run where they currently sit.

2. unresolved_count: 0 is stronger than what could be verified. Of 1,345 PRESERVED_*
entries, 1,322 resolve to a file present in the head. The other 23 point at paths in neither tree,
all demo videos and screenshots under Feature list stuff/. No real risk, but the count reads as
covering the whole set.

3. Two new name collisions. decision_artifact and privacy_control now exist in both
harness/lib and harness/tools, taking the total from 313 to 314. This matters because
tests/conftest.py resolves tools ahead of lib, so the suite imports the CLI wrapper instead of
the implementation. That ordering bug is fixed in #44, but the collisions are worth avoiding.

Test impact: 255 currently-failing tests were run against this head and against the base under
identical conditions. 7 are genuinely fixed, 0 regressed, 242 remain red. The 7 include
test_capability_capsule_definition_assembly__duplicate_identity_version, which traces to
51a0b00.