Phase log 18: 2026-08-24 - Card sizing, repo cleanup, and the CLAUDE.md drift audit

Date: 2026-08-23 → 2026-08-24 Branch: main, at 0088c90 QOL Improvements. Nothing committed this session.

Goal

Started narrow: make the card-size setting change card height as well as grid column width, stop text clipping without an ellipsis, and fix the copy-button footer changing size and position from card to card.

Then widened twice, at the user's direction: first to a critique of that change, then - after a clarification - to an all-time whole-repository diagnosis. The diagnosis was interrupted twice (once by the usage limit, once by the user), and only one of its eight reports survived. That report is the reason the second half of this session became a documentation-accuracy fix.

Current Progress

1. Card sizing - done and verified

src/css/cards.css now derives every card dimension from --cs-* tokens selected by #card-grid[data-card-size], which applyGridSettings() stamps in src/ts/state.ts. One setting moves width and height together.

Collapsed cards have a fixed height; .card-head takes the slack (flex: 1 1 auto) so the footer sits on the bottom edge. Expanded cards are height: auto and every clamp is lifted. #card-grid gained align-items: start so an expanded neighbour cannot stretch a collapsed card.

Verification: npx vitest run → 671 passed / 27 files; npm run typecheck clean; npx vite build clean. Nine new tests in tests/render.test.ts, each verified by reverting its fix and watching it fail.

2. Abandoned worktree deleted - 1.9 GB reclaimed

.claude/worktrees/agent-a3723c0e15ae0afb5 was registered under the pre-rename path ~/Code/APIV/..., so its gitlink was already broken (git status inside it returned "not a git repository"). phase log 1 §Next Steps had prescribed this cleanup and recorded it as blocked by the sandbox classifier.

Verified safe before deleting: its commit 6284b35 ("Initial Commit at second phase", May 2026) is an ancestor of main, branch worktree-agent-a3723c0e15ae0afb5 still resolves, and its six locally-modified files were Phase-3-era edits long superseded (its lib.rs was 654 lines against main's 958). Those six are archived at scratchpad/orphan-worktree-modified-files.tgz regardless.

Repo is now 9.4 MB excluding node_modules/target.

Not done from phase log 1's cleanup list: git branch -D worktree-agent-a3723c0e15ae0afb5 (kept deliberately as deletion insurance) and rm -f architecture.dot. Both still pending a decision.

3. CLAUDE.md drift fixed

See "The bugs that mattered" below for what was wrong. Nine corrections applied, plus a new "Keeping this file true" section stating the rules that would have prevented them.

The bugs that mattered

Ranked by consequence. The first is the worst thing found this session, and it is not in the code.

1. Five consecutive documents asserted work was undone that was already committed

CLAUDE.md's Immediate Next Steps and Handoffs 2, 3, 4, 5 and 6 all said "Commit Phase 11 - still uncommitted, now underneath 0088c90". It was committed in de212a3 "Fixed Bugs" at 14:37:47, thirty minutes before 0088c90 at 15:07:37. de212a3 contains all 24 Phase-11 test files.

Nobody checked git log. Each session inherited the line and copied it forward, and it became more credible with each repetition. What that cost: every session since Phase 11 opened with a false top-priority item, and the real uncommitted work - 8,036 lines across 38 untracked files, the whole of Phases 13–15 - was never described accurately in its place.

This is the same failure mode the project already documented in Phase 6, where a wrong claim about danger_accept_invalid_certs "hid a real bug". The documentation is load-bearing here because it is injected into every session.

2. The Architecture section described a codebase shape three phases dead

### File Layout listed src/vault.ts (~1800 lines) and src/vault.css. Neither has existed since Phase 3 split them into 24 modules under src/ts/ and 15 stylesheets under src/css/. A reader trusting it would look for files that are not there and miss vault-core/, envv-server/ and envv-cli/ entirely.

3. Type unions documented at a third of their real size

ProjectType was listed as 3 variants; there are 11. ChunkType as 7; there are 29. Both grew across Phases 12–15 without the doc moving.

Worse, the ProjectType line contradicted the Phase 12 section ninety lines above it, which correctly lists all 11. Two sections of one file disagreeing is a live bug in the docs, and the nearer one was wrong.

4. The Rust command list documented 9 of 36

Missing: the entire RBAC surface (users, classes, tokens - 17 commands), all three LAN commands, get_expiring and get_audit_log. Anyone reading CLAUDE.md to learn the Tauri API would conclude multi-user and Open-to-LAN do not exist.

5. The vault_audit schema omitted the hash chain

Documented as five columns. The real table gains entry_hash, prev_hash and actor via idempotent ALTER TABLE in init_schema(). Those three columns are the tamper-evidence that envv scan --verify checks - the doc omitted the feature, not just the columns. A third table, vault_meta, went unmentioned.

6. A Platform note contradicted by its own Phase 15 section

## Platform said GDK_BACKEND=x11 is set unconditionally in run(). Phase 15, 700 lines earlier in the same file, correctly says it is set only when the session is not Wayland - and configure_linux_webkit() in src-tauri/src/lib.rs:835 confirms Phase 15 is right. Following the Platform note would mean forcing XWayland and reintroducing the fractional-scaling bug Phase 15 fixed.

Also in that section: the project path was ~/Code/ENVV; the repo is ~/Code/EnvVault.

7. Three different test counts in one file

563 (Phase 11 row), 651 (Phase 12 row), 662 (Testing header and the npm test comment). The real number is 671. Recording a number in four places is four chances to disagree; it now lives only in the Testing section.

Decisions made

• Fixed collapsed card height, not a min-height. "Change the height" could mean either, but only a fixed height puts the footer at a constant offset - which was the third requirement. Cards with little content carry slack in the head rather than shrinking.
• Clamp to whole rows, never a fraction. Badge rows, tags and the project row clamp to an exact multiple of their row height, so a clamp never slices a badge in half. Ellipsis is used wherever the overflowing thing is text.
• Test counts recorded once. The Testing section is the only place. Phase rows describe what a phase did, not how many tests existed when it ended.
• Lists that mirror enumerable things now carry the enumerating command. grep -c '#\[tauri::command\]' sits beside the command list; src/ts/types.ts beside the unions. A reader can check staleness in one step.
• Kept the orphan worktree's branch. Deleting 1.9 GB of working tree is reversible from the branch; deleting the branch is not. phase log 1 asked for both; I did the expensive half only.
• Did not restructure the three-document contract. CLAUDE.md and CLAUDE.original.md both violate it (see Next Steps). Splitting them is a larger, riskier edit than the factual corrections asked for, and it should be a deliberate decision rather than a side effect.

What Worked

• Reverting each fix to verify its test. Four of nine new tests failed on the first revert pass; a fifth only failed after I reworded a CSS comment, because the regex was matching the literal min-width:0 inside the comment rather than the declaration. That test had been passing for the wrong reason. This technique keeps earning its cost.
• Verifying doc claims against git log instead of reading more docs. The Phase 11 finding took one command. Five sessions of reading had not found it.
• Checking a destructive delete three ways first - commit reachability, branch existence, and a file-by-file hash comparison against the commit - then archiving the six differing files anyway.

What Didn't Work

• Eight parallel Opus subagents over a 40k-line repo. Exhausted the session limit before any of them finished. Every report was lost because they were told to write at the end.
• Relaunching all five at once after the reset. Better, but still too wide; the user stopped it. Of five, only the docs/tests agent produced anything substantive - and only because that run wrote incrementally.
• Waiting after the usage-limit reset instead of resuming. The user had to say "you could have restarted". The reset notice ends the checkpoint constraint; it is not a standing stop.
• Regexing CSS source as a test. Five of the nine new tests do this because jsdom applies no stylesheets. They are a brittle proxy: they match comments, they anchor to the first matching rule and miss later cascade overrides, and one (--cs-card-h ordering) depends on declaration order in the file. They are better than nothing and worse than a real layout check.

Next Steps

1. Commit the tree. 28 modified files (+2,335/−646) and 38 untracked files (8,036 lines) - Phases 13–15. This is the largest unversioned surface the project has ever carried. A commit plan was requested from the diagnosis agent but that section was never written; it needs redoing. The user commits, not Claude - hand over a pwsh one-liner.
2. Finish the diagnosis. Eight areas were scoped; one produced a partial report (scratchpad/diag-8-docs-tests-ci.md, Sections A and B only - C/D/E were never written). Not started: Rust core/Tauri, envv-server, envv-cli redaction, frontend invariants 1–9, chunks/exporters TS↔Rust drift, authz/remote/LAN, CSS/UX/a11y. Run these in waves of two or three, on Sonnet for the inventory work, and require incremental writes to disk.
3. CLAUDE.md is not tracked by git. .gitignore:30 ignores .claude/, and CLAUDE.md, CLAUDE.original.md and every Handoff-N.md are untracked. The project's only memory has no history, no diff, and no backup - which is part of why drift went unnoticed for five sessions. Decide whether to track them.
4. The three-document contract is not being followed. CLAUDE.md carries a ## Phase History and a ## Handoff - Post-Phase 6 Session (both session-log material it says it must not hold). ~74% of CLAUDE.original.md's 953 lines are a dated chronological Session History rather than pitfalls grouped by area, and it has no Phase 15 entry at all. Three overlapping session logs exist and none is authoritative.
5. Still-open items confirmed by the audit (all pre-existing):
   • chunks/env-link.ts and chunks/edit-modal.ts at 0% coverage.
   • The local vault key stays resident in VaultState for a whole remote session - no decision since Phase 12.
   • The CLI cannot pin a certificate (no --ca-cert, no custom verifier in access.rs), so a self-signed envv-server is unreachable from envv.
   • Experimental exporters (apache/haproxy/ansible/postgres) are UI-only.
   • No tests for the Phase 15 icon validator, readIconField/setIconField round-trip, the enrich signature table, or the is_blank secretType case.
6. Five manual smoke tests, still never run - listed in CLAUDE.md §Immediate Next Steps #2 since Phase 11. Only a human can do these.
7. Decide on architecture.dot and the orphan branch (phase log 1's remaining cleanup).

Notes for whoever picks this up

The card work is finished and green; do not re-litigate it. The interesting material this session is documentary, not functional.

The lesson worth carrying: this project's docs are injected into every session, so a false line in them is not an inaccuracy, it is a recurring instruction. "Commit Phase 11" survived five sessions and cost each of them a wrong first priority. Before repeating any claim you inherited from CLAUDE.md or a handoff - especially one about the state of the repo - spend the one command it takes to check. CLAUDE.md now has a "Keeping this file true" section saying so.

If you resume the diagnosis, keep the surviving report (scratchpad/diag-8-docs-tests-ci.md) - Sections A and B are verified work and should not be redone. Its promise ledger is the single most useful artifact produced this session.