Standards batch shipped to a commit; push blocked at the sandbox boundary (2026-06-12)

Picked up the LLM-Knowledge-Base redesign thread to push it to GitHub. The redesign’s already-committed work was one commit ahead of origin/main (the 06-12 journal entries, f7eb64c); the open piece was the “review before they ship” batch from the prior session, still sitting uncommitted in the working tree. Reviewed it, committed it, and hit a hard boundary on the push.

Decisions

The review batch is sound — committed as bd5adc9 (8 files, +326/−194). Verified before committing rather than trusting the prior session’s summary:

• Writing Standards.md reorder is content-preserving. Headings confirmed: the three High-Signal sections (Decision Writing, Wiki Pages, Condensed Pages) moved to the top as the primary standard; the May material moved below a new ## Supporting Detail — The May Standard (subordinate) divider, original order intact. Word count 5,929 → 5,982 (+53 = the divider plus fixed self-references). Nothing deleted.
• GROK.md — authority order now stated explicitly (High-Signal sections lead; May material subordinate, loses conflicts). Consistent with the reorder.
• PRD-Site-Aesthetic-Refresh.md and 2026-05-30-site-full-send-wnab-aesthetic.md — status: superseded with superseded-by pointers to the Living Atlas redesign; Lora + JetBrains Mono + amethyst noted as the surviving tokens.
• Four backfilled decisions/ notes (engine Quartz-over-Astro, Living Atlas design, ledger demotion H7, High-Signal standards), all type: decision so the Command Center log picks them up.

Wedge reviewed locally before sign-off. Asked to see diffs first, reviewed in Obsidian, then authorised “commit all.” No blind commit.

Constraints discovered (ops — these will recur)

Pushes can’t originate from the Cowork sandbox. git push to the HTTPS remote fails could not read Username for 'https://github.com' — the sandbox has no GitHub credentials and can’t prompt. This is why the prior session’s pushes/merges were done host-side too. Standing rule: Cowork commits; Wedge pushes from /Users/n1/Projects/llm-knowledge-base where his git auth lives.

The sandbox mount forbids unlink inside .git. Every git command that touches the index leaves an index.lock it can’t delete, which then blocks the next command (File exists). Workaround that works: mv the lock out of the way (rename is permitted where unlink isn’t), then run the target git op as the only git call in that shell invocation — no preceding git status, or it re-poisons the lock. Commit succeeded this way. Cosmetic residue left behind in .git/: index.lock.stale, index.lock.old2, HEAD.lock.old (0-byte; Wedge to rm host-side).

Open / outstanding

• Push still pending. Two commits queued on main ahead of origin: bd5adc9 (standards) + f7eb64c (journal). Wedge runs git push origin main to publish to the live Pages site. (This entry will be a third commit once written.)
• Dangling-link risk, unresolved. The new superseded-by pointers reference journal/2026-06-11-living-atlas-and-decision-writing and PRDs/atlas/. The prior session flagged that the Living Atlas PRD and the two parent redesign proposals cited in PRDs/atlas/README.md may not exist in the repo. Confirm these resolve before/after the rebuild, or the live site ships broken links.
• Inherited threads still open (from the redesign-review session, not closed here): Living Atlas PRD reconstruction handed to Grok/Composer; the enableSPA test is still unmeasured though the engine decision leans on it; auto-catalog PRD scheduled “after Atlas ships”; ~105MB worktree duplicates + four merged branches left un-pruned (never approved).