docs: add review-proven failure-path & async-state rules to AGENTS.md (#7061)

Mining the last 25 PRs' review threads (45 substantive findings, 11
reviewed PRs, avg **4.8 review rounds** each) shows **53% of findings
are repeats** of five clusters: swallowed failures, stale-async-state
races, tests that don't bind the production seam, unbounded
resources/retry loops, and non-atomic multi-step persistence. PR #6956
alone burned 4 rounds converging on one of these classes.

A second, independent mining pass over **71 agent-review rooms (303
findings, Aug 18–29)** confirmed the same clusters and added outcome
data — how often authors actually fix each finding class once flagged:
test-seam binding and unbounded-resource findings **100%**, swallowed
errors **90%**, stale-state races **70%**. It also surfaced two clusters
the GitHub-thread pass under-sampled: **assistive-semantics defects**
(44 findings, second-largest cluster) and **input-modality divergence**
(27 findings), now rules 7–8.

This PR distills those clusters into eight imperative rules in AGENTS.md
so agents apply them **before writing code**, adds one
client-consumption invariant to ARCHITECTURE.md §5, and places the
test-quality rule in TESTING.md (per the team decision that testing docs
are the canonical guide for review standards), cross-referenced from
AGENTS.md. Each rule cites the PRs where it was litigated. Raw mining
data: `reviews.jsonl` / `comments.jsonl` +
`backfill/buzz-review-findings.jsonl` (review-mining artifacts, not
committed).

No code changes. CLAUDE.md is a symlink to AGENTS.md and picks this up
automatically.

🤖 Drafted by Jude's agent from automated mining of this repo's last 25
PRs' review threads and 71 agent-review rooms; every rule cites the PRs
where it was litigated. Jude reviews and owns the result. Mining method
+ raw cluster data available on request.

---------

Signed-off-by: Jude Edwards <judeedwards@squareup.com>

AGENTS.md

@@ -151,6 +151,91 @@ Additional rules:
 
 ---
 
+## Review-Proven Rules
+
+These rules distill the recurring findings from the last 25 PRs' review
+threads — 53% of substantive review findings were repeats of the clusters
+below, and reviewed PRs averaged ~5 review rounds. A second, independent
+mining pass over 71 agent-review rooms (303 findings, Aug 18–29) confirmed
+the same clusters and measured how often authors actually fix each class
+once flagged: test-seam binding and unbounded-resource findings were fixed
+**100%** of the time, swallowed-error findings **90%**, stale-state races
+**70%** — these are not style opinions, they are defects authors agree
+with on sight. Apply the rules **before writing code**; each cites the
+PRs where reviewers litigated it.
+
+1. **Every caught failure must leave a durable retry record or propagate.**
+   Never catch-log-and-return-success (opt-out revocation permanently
+   abandoned, PR #6269), never convert a terminal failure into an
+   authoritative success/empty result (cold-history `error` → `success`
+   with `[]`, PR #7013), and never delete the durable journal an operation
+   depends on before its retry has actually succeeded (PR #6269). If a
+   partial failure can orphan committed state (installations, endpoints),
+   schedule its cleanup/renewal durably (PRs #6269, #6996, #7013).
+
+2. **Fence async results by generation; clear derived metadata on every
+   removal path.** A completing in-flight probe or fetch must verify it is
+   still the newest before writing its result (stale login-shell probe
+   recached a false-negative PATH, PR #6904). Provenance/ownership metadata
+   attached to synthetic state must be updated or cleared on *all* paths
+   that remove or refresh that state — typed deletion, toolbar removal,
+   profile/name refresh; enumerate the paths and test each (PR #6956 burned
+   4 rounds on this one class). Backfill and live subscriptions must
+   overlap — a gap between a finite history REQ and the live subscription
+   silently drops events (PR #3995); a retired chunk must not keep a stale
+   scope fence (PR #6996). (PRs #3995, #6904, #6956, #6996)
+
+3. **Regression tests must bind the production seam and be falsifiable.**
+   See "Review-Proven Test Standards" in [TESTING.md](TESTING.md) for the
+   full rule — in short: a guard whose removal doesn't fail any test
+   protects nothing; bind regression tests to the production code path,
+   not test-only helpers. (PRs #6807, #6980, #6996, #7013)
+
+4. **Bound every resource, loop, and process tree.** Cap captured
+   output (unbounded discovery temp files exhausted disk and overran the
+   deadline, PR #6904). Containment failures are errors, not warnings — a
+   tolerated Job Object creation failure or a `setsid` escape leaks whole
+   process trees (PR #6904). Retry/re-subscribe loops need backoff and a
+   terminal state: a persistent failure must not self-amplify into an
+   unbounded refresh loop (PR #6996), and check zero-delay edge cases
+   (`remainingMs()==0` selected the wrong fallback window, PR #6996).
+   (PRs #6904, #6996)
+
+5. **One user action = one atomic persist.** Implementing a single user
+   commit as N independent durable writes leaves torn state on partial
+   failure (theme "Set" as three independent notifier persists, PR #6944;
+   relay-commit vs. local-save recovery gap, PR #6269). Persist one
+   snapshot, or order the writes so every prefix is consistent and the
+   remainder is durably retried per rule 1. (PRs #6269, #6944)
+
+6. **A guard that hides the only recovery affordance is a functional
+   failure.** Before adding a visibility predicate or state fence, ask:
+   if the state it assumes goes wrong, does the user still have a way
+   back? A fence that permanently suppresses "jump to latest" after a
+   bounded correction fails strands the user silently — two reviewers
+   flagged this independently (PR #6807).
+
+7. **Audit assistive semantics on every new visual component.** The
+   agent-review lanes flagged accessibility defects on 44 findings across
+   the Aug 18–29 window — the second-largest cluster — and authors fixed
+   the concrete ones (duplicate VoiceOver stops on native controls,
+   actionable labels owned by two widgets at once, PR #6680; missing or
+   decorative-leaking semantics on new UI, PRs #6611, #6702, #6885, #6905,
+   #6908). New UI ships with: one owner per actionable label, no duplicate
+   screen-reader stops, and explicit semantics for every interactive
+   element. (PRs #6611, #6680, #6702, #6885, #6905, #6908, #6980)
+
+8. **Every input modality is a first-class seam.** Keyboard, pointer, and
+   hotkey paths must not silently diverge: `Shift+Space` treated as plain
+   `Space` because the guard omitted `shiftKey` (PR #6862), keyboard
+   ownership not released on blur, modifier keys dropped on the non-mouse
+   path (PRs #5958, #6793, #6860, #6908, #7006). When adding an input
+   handler, enumerate the modalities that can reach it and test the
+   non-primary ones — that's where the defects were. (PRs #5958, #5972,
+   #6793, #6860, #6862, #6908, #7006)
+
+---
+
 ## Key Patterns
 
 **Nostr-first HTTP surface**: Buzz's primary API is NIP-29 over WebSocket. The relay also exposes a narrow HTTP surface: NIP-11/NIP-05 metadata, `POST /events`, `POST /query`, `POST /count`, workflow webhooks at `/hooks/{id}`, Blossom media, git smart HTTP, git policy hooks, and health probes. These HTTP paths all preserve the same host-derived community boundary.

ARCHITECTURE.md

@@ -325,6 +325,15 @@ This prevents a race where a non-member receives live fan-out events from a priv
 
 After registering, the REQ handler queries Postgres for stored events matching the filters (up to 500 per filter, hard cap). These are sent as `["EVENT", sub_id, event]` frames before `["EOSE", sub_id]`. New events arriving after EOSE are delivered via the fan-out path.
 
+**Client consumption invariant.** A client rebuilding channel state must
+open its live subscription before (or overlapping) the finite history
+REQ — a gap between the last backfill page and live delivery silently
+drops events and rebuilds stale state (PR #3995). When the relay sends a
+terminal CLOSED, the subscription is removed server-side; any client-side
+ownership tied to it (chunk/scope fences) must be released in the same
+step, or live delivery stops permanently while the client believes it is
+subscribed (PR #6996).
+
 ---
 
 ## 6. Crate Reference

TESTING.md

@@ -16,6 +16,21 @@ just test               # unit + integration (starts Docker if needed)
 cargo test -p buzz-test-client -- --ignored
 ```
 
+### Review-Proven Test Standards
+
+Mined from the last 25 PRs' review threads (see Review-Proven Rules in
+[AGENTS.md](AGENTS.md)); this is the test-quality rule reviewers litigated
+most:
+
+**Regression tests must bind the production seam and be falsifiable.**
+A guard whose removal doesn't fail any test protects nothing — mutations
+survived the full mobile suite twice (PRs #6996, #7013). Don't bind a
+regression test to a test-only helper instead of the production code
+path (PR #7013). Give pure predicates a table test over the full input
+combination space (PR #6807). Scope Playwright locators — unscoped
+`getByText` in a required smoke test is a strict-mode flake (PR #6980).
+(PRs #6807, #6980, #6996, #7013)
+
 ---
 
 ## Live Local Relay