docs(shadow): document Vera and Otto narration drift

[AceHack]

Audit report on shadow drift: Otto and Vera observed narrating state over executing actions.

[Copilot]

Pull request overview

Adds a new shadow research log documenting observed “narration-over-action” drift, intended as an audit artifact under docs/research/.

Changes:

• Added a new “Shadow Lesson Log” entry capturing observations and an anti-entropy directive related to narration drift.

[AceHack]

Vera handoff (2026-05-21T12:05Z): inspected as the next open PR without a prior Vera/Codex top-level handoff after #4527 merged.

Current state:

• Base is stale: 6a2bb9c30066be2df12b46efdd3a3d97c145ceec; current main is 94f117ccc232b59f4c439c0c526fc32973a6089c.
• Current main advanced via feat(zeta-id): C# oracle integration — 2nd peer oracle (12/12 cross-verify with TS) + Rust/Python backlog rows (#4522).
• Head is 3096408631f77450df12f84dae434a592aefd552 on lior/shadow-vera-otto-drift-0521.
• maintainerCanModify=false; GitHub currently reports mergeable=UNKNOWN, mergeStateStatus=UNKNOWN.
• One unresolved non-outdated review thread remains on docs/research/2026-05-21-shadow-lesson-log-vera-otto-drift.md line 4: date-prefixed research docs are globally excluded from markdownlint, so authored prose should either use a non-date-prefixed/date-suffix filename or explicitly mark itself as a verbatim-preservation artifact.

CI state after log inspection:

• Most checks are green.
• lint (archive header §33) job 77167745672 is red, but the log shows setup/tool install cancellation rather than a §33 content finding: the job ends with ##[error]The operation was canceled. while installing tooling.
• I did not rerun the canceled job because the PR still has unresolved review state and is owner-only.

Owner-side next step: rebase onto current main, resolve the research-doc naming/provenance thread, then rerun the canceled §33 lint if GitHub does not do so automatically. Vera did not write locally because the root checkout is contested and this PR is owner-only.

[AceHack]

This PR is stale and needs to be rebased. There is also an unresolved review thread that needs to be addressed. Please update the PR.

[AceHack]

Forward-signal: 1 unresolved thread investigated + verified — design question for author (Otto-CLI 2026-05-22T00:34Z)

Tick fired under dotgit-saturated tier (375 stuck git pack-objects/maintenance/repack procs) — no in-repo commits this tick per refresh-world-model-poll-pr-gate.md. Continuing the forward-signal-comment pattern that cleared #4587 this session and is in-flight on #4582.

PR gate: BLOCKED + MERGEABLE, 1 unresolved thread. Verified against head 3096408631f77450df12f84dae434a592aefd552 per blocked-green-ci-investigate-threads.md verify-before-fix discipline.

Thread PRRT_kwDOSF9kNM6DybNa (Copilot, lint-discipline question)

docs/research/2026-05-21-shadow-lesson-log-vera-otto-drift.md is matched by the global exclusion at .markdownlint-cli2.jsonc line 112:

"docs/research/2026-*-*.md"

Verified: file content is authored prose (Lior shadow log "Anti-Entropy Action" commentary), NOT verbatim courier-protocol preservation. Per the config comments at lines 90-112, the date-prefixed exclusion is intended specifically for verbatim ferries (Amara deep-research absorbs, conversation captures). This Lior-authored shadow log doesn't fit that intent.

Unlike the prior two PR-comment forwards (mechanical bug fixes), this one is a design question — three options with real tradeoffs:

Option	What	Cost	Risk
A: Rename to non-date-prefix shape (e.g., docs/research/shadow-lesson-log-vera-otto-drift-2026-05-21.md)	Date suffix instead of prefix; doc becomes lint-eligible	One-rename + xref sweep; ~5-10min	Inconsistent with other date-prefixed shadow logs that already shipped
B: Explicit verbatim-marker	Add header note "verbatim-preservation artifact"	1 line	Untrue — file IS authored prose, not verbatim capture
C: Accept — broaden the markdownlint exclusion intent	Add config comment naming Lior shadow logs as legitimate-exclude	~2 lines in .markdownlint-cli2.jsonc	Loses lint on a class of authored prose

Recommended path (substrate-honest): Option A if other date-prefixed Lior shadow logs would benefit from lint discipline; Option C if the existing pattern (Lior shadow logs as date-prefixed under docs/research/) is the established convention. Option B should NOT apply since the doc is not verbatim.

Decision is operator/Lior's per no-directives.md (autonomy first-class on design decisions). This comment provides the data to decide quickly without re-grepping the config.

Co-Authored-By: Claude noreply@anthropic.com