auto-fix batch claude/friendly-maxwell-0olke (2026-04-29)

[intendednull]

2 small-scope fixes from open issue queue + 1 structural follow-up. Branch rebased onto current main (which already has #484's skill rewrite). Force-pushed clean — mergeable_state should now be clean, full CI re-runs.

Fixes

• Fixes [ARCH-11] crates/state/src/tests.rs is 3,726 LOC — single-file monolith, 44% of the crate #263 — split crates/state/src/tests.rs (4,714 LOC monolith — issue cited 3,726, drift) into 6 per-concern files under crates/state/src/tests/. Wired via #[path] matching crates/client/src/lib.rs:49-55. 117 #[test] fns preserved verbatim; full state-crate suite 227 → 227 green. Largest bucket (materialize.rs tests, 1,916 LOC) now smaller than materialize.rs source (1,625 LOC). Sixth bucket tests/stress.rs added beyond the issue's 5 because the perf cases didn't fit the per-concern split cleanly. Tradeoff: each bucket carries duplicated genesis_kind / test_dag / do_emit helpers (~30 LOC each) under #![allow(unused_imports, dead_code)]; runner-up of factoring into tests/common.rs rejected as new abstraction beyond the mechanical-split scope.
• Fixes [GEN-12] tokio::time::sleep(50ms) used to "wait for actor propagation" in ~10 ClientHandle methods #271 — drop tokio::time::sleep(Duration::from_millis(50)) "wait for actor propagation" patterns in crates/client/src/lib.rs. 10/10 sites converted: 6 × bucket-DROP (StateActor's FIFO mailbox already orders the next call — send_message_and_read_back, switch_server_updates_event_state ×2, generate_invite_grants_send_permission_to_recipient, mutate_channel_mute_emits_event_and_flips_stats, mutate_channel_mute_toggle_off_clears_set); 4 × bucket-POLL (derived-actor lag — Notify → spawn → UpdateCache — presence_self_override_round_trip, presence_reachable_peer_defaults_to_here, presence_queued_then_gone_after_threshold ×2; new await_view(probe) helper, 5s timeout, mirroring wait_for_message/wait_until in multi_peer_sync.rs); 0 × bucket-ASK (every site already used state::mutate(...).await, which is ask under the hood — sleeps were cargo cult); 0 × bucket-WATCH. Stale-issue note: issue body claimed sleeps were "outside #[cfg(test)]"; they were all inside the #[cfg(test)] mod tests block (lines 1255-2082) after a recent test-file refactor. Fix pattern still applies. Diff: 1 file, +63/-19. cargo test -p willow-client --lib 303 passed before + after; presence tests ran 3× clean.

Already-Fixed / Parked

• [TD-02] convert_case linked at three incompatible versions (0.6, 0.7, 0.10) #251 [TD-02] convert_case 3 versions — same shape as [TD-03] getrandom linked in four versions (0.2 / 0.3 / 0.4 x2) — feature-unification risk for WASM #252/[TD-03 follow-up] getrandom 3-way split is structural — driven by aes-gcm 0.10, ahash, and iroh 0.98 #481. Three majors (0.6.0, 0.7.1, 0.10.0) pinned by transitive crates we don't own:

  • 0.6.0 ← server_fn_macro 0.7.8 + config 0.15.22 (Leptos 0.7 internals)
  • 0.7.1 ← leptos_macro 0.7.9 + reactive_stores_macro 0.1.8 (Leptos 0.7 disagrees with itself)
  • 0.10.0 ← derive_more-impl 2.1.1 (entire iroh stack: iroh, iroh-base, iroh-blobs, iroh-gossip, iroh-relay, iroh-tickets, n0-future, n0-watcher, netwatch, portmapper, noq, irpc)

  cargo update -p convert_case is a no-op — all already at latest in their semver range. Workspace [patch] can't bridge ^0.6 ↔ ^0.7 ↔ ^0.10 without lying about semver, which would break derive_more's actual API usage. Real fix requires three separate upstream events (Leptos 0.8 internal alignment, derive_more 3.x ecosystem catch-up). Filed follow-up [TD-02 follow-up] convert_case 3-way split is structural — pinned by Leptos internals + derive_more #485 documenting the holdout chain. [TD-02] convert_case linked at three incompatible versions (0.6, 0.7, 0.10) #251 closed not_planned (audit's "obvious fix" premise was structurally moot; not just a missed PR).

  This is the 3rd structural-deps follow-up in a row (rand [DEP-02] Three concurrent rand major versions + RUSTSEC-2026-0097 across all of them #246, getrandom [TD-03 follow-up] getrandom 3-way split is structural — driven by aes-gcm 0.10, ahash, and iroh 0.98 #481, convert_case [TD-02 follow-up] convert_case 3-way split is structural — pinned by Leptos internals + derive_more #485). Skill update below codifies an "upstream-domino meta-tracker" path so the next run files a single tracker instead of accumulating individual TD-NN follow-ups.

Skill Evolution

Single skill commit on this branch: a8581be — two lessons folded into the new SKILL.md structure landed via #484:

• Implementer step 2 (research) — re-grep cited line numbers / LOC counts at HEAD before working from issue-body literal positions. [ARCH-11] crates/state/src/tests.rs is 3,726 LOC — single-file monolith, 44% of the crate #263 cited 3,726 LOC (actual 4,714); [GEN-12] tokio::time::sleep(50ms) used to "wait for actor propagation" in ~10 ClientHandle methods #271 cited "outside #[cfg(test)]" but the sleeps had since shifted INSIDE a test module. Drift is the norm.
• New Implementer step 12 — structural-deps follow-up family path. On the 3rd structural-deps follow-up, file/update an upstream-domino meta-tracker issue rather than another standalone TD-NN follow-up. Lists the holdout crates, upstream releases that would unblock each.

Earlier rev of this branch also tried to promote "mechanical splits / file moves" to a top pick in Core Loop §2; reviewer pushed back on the prioritization framing (would skew the agent toward cosmetic moves over real fixes), so it was dropped on rebase.

Lessons Learned

• Sub-PR vs direct-commit pattern. This run executed under the OLD SKILL.md (worktrees + sub-PRs), then the rebase-onto-main collapsed it into the NEW shape (commits direct on master, no sub-PRs). End state is identical — the new shape is just cleaner from the start. Future runs that begin with the new SKILL.md will skip the sub-PR overhead entirely.
• Pre-dispatch sync (git fetch + reset --hard origin/) held clean across 3 dispatches.
• Webhook-informational rule held. Coordinator received 2 sub-PR open + 2 sub-PR merged webhooks (refactor(state): split tests.rs monolith into per-concern files #486, kill 50ms sleep waits in client lib.rs tests #487); never broke flow on any of them.
• Caveman-mode GH comms kept master + sub-PR bodies tight. Code blocks stayed normal.
• Complexity gate caught a 3rd Category 3. [TD-02] convert_case linked at three incompatible versions (0.6, 0.7, 0.10) #251's "obvious fix — will be auto-PR'd" framing was structurally moot, just like [TD-03] getrandom linked in four versions (0.2 / 0.3 / 0.4 x2) — feature-unification risk for WASM #252. The implementer correctly investigated cargo tree first instead of pushing a half-fix.
• Stale-issue framing is now confirmed common. Hence the new "re-grep at HEAD" addition to step 2.
• Cross-session skill-commit interaction. Original rebase tripped a conflict in SKILL.md because main had auto-fix batch claude/friendly-maxwell-bVlpW (2026-04-28) #484's full rewrite landed mid-flight. Recovered cleanly via hard-reset + replay code commits + re-author skill edits onto new structure. Future runs that span a SKILL-rewrite merge should expect this and prefer reset-and-replay over cherry-pick-and-resolve when the upstream structure has drifted.
• 3 dispatches vs prior run's 5. Both fine. Cap is 10 but quality > quantity.

Test plan

• Master-PR CI runs full just check (fmt + clippy + test + wasm + browser + Playwright).
• Manual smoke: just dev + agent-browser navigation, focused on:
  • Server creation + channel creation → confirms crates/state materialization still passes through the split test buckets in production code paths (no test-file regression risk for production, but the #[path] wiring in lib.rs is the only behavioural surface that could break).
  • Send message → channel message → switch server → presence — confirms the bucket-DROP and bucket-POLL conversions in lib.rs tests didn't drift in client-handle semantics. Tests pass; smoke confirms the same flows still work end-to-end.

---

Generated by Claude Code

[intendednull]

resolve conflicts