Persist the identity cache only when it changes

Summary

SecureIdentityStateManager.upsertCryptographicIdentity rewrote the whole identity cache on every verified announce, even a byte-identical re-announce. Each rewrite is a JSON encode, an AES-GCM seal, and a keychain delete+add, and it grows with every peer ever seen.

Measured (M5 Max, in-memory keychain): 1.6 ms per announce at 500 identities, 6.2 ms at 2,000 and 33 ms at 10,000, plus 3-8 ms of keychain work on a device. Main-thread isVerified reads waited behind the save barrier. A batch of 16 vouches did 16 synchronous full rewrites on the main thread: 101 ms at 2,000 identities on the Mac, roughly 0.35-0.65 s on an older iPhone.

This PR writes only when something changes, batches vouches into one write, and indexes short-ID lookups. It also fixes two persistence bugs the change exposed.

Retention (bounding the cache size) is not in this PR. It changes TOFU and privacy behavior and has its own stacked draft PR for that decision.

Changes

Fixes the change surfaced:

• A single failed keychain write could lose the whole identity cache. KeychainManager.saveData deletes, then adds. If the add failed, the item was gone, but pendingSave was still cleared, so forceSave at termination did nothing. Today the next announce papered over this by accident. With the change gate it wouldn't, so a failed write now stays pending and is retried:
  • on the next real change;
  • on forceSave / deinit;
  • at most every 30 s on unchanged upserts.
• bindAuthenticatedSigningPublicKey changed the crypto identity's signing key without saving it. It was only persisted by the unconditional upsert that followed. It now saves what it changes.

Measured

M5 Max, in-memory keychain (securityd cost excluded), C = 2,000 identities unless noted. Before = origin/main with the same benchmarks.

On an older iPhone (CPU 3-5× slower, 3-8 ms per keychain write), an unchanged announce at 2,000 identities goes from 24-44 ms to microseconds, and a vouch batch from 355-640 ms to about 25-45 ms. A first contact still writes once, as before: that is a real change.

Behavior

• Unchanged: persisted state and the stored format. Only no-op rewrites are skipped; no new fields.
• Unchanged: the BLE/Nostr wire format and TOFU behavior. A pinned-key mismatch still refuses with no write.
• Untouched: app lifecycle, Tor, AppRuntime, timers, Tasks, and background hooks. The only lifecycle-reachable change is that forceSave() at termination now retries a previously failed write.

Tests

• New, each mutation-checked:
  • 7 SecureIdentityStateManagerTests: no rewrite on unchanged re-announce, rewrite on real changes (with reload), peer-state sequence writes, bind persistence, forceSave retry, retry on next announce, and retry backoff;
  • BLEAnnounceIdentityPersistenceTests (real handler, real signatures);
  • 4 vouch-batch tests and handleVouchPayload_recordsAcceptedAttestationsInOneCall;
  • a prefix-index equivalence test against the old linear scan;
  • 3 benchmarks with floors.
• SwiftPM: 2,166 Swift Testing tests, 0 failures, on this branch's tree. The full branch with retention passed 2,175 + 247, twice.
• Builds, lint, perf gate: macOS build is clean. SwiftLint and Periphery have no new findings versus main. The perf gate passes on this branch (19/19 floors).
• iOS simulator: 2,396/2,396 passed on this branch (dedicated iPhone 17 Pro Max simulator, iOS 27.0, serial as in CI).

Merge notes

• Conflicts: the only conflict with #1809 and the announce PR is bitchatTests/Performance/perf-floors.json, because each PR adds its own floors and reference numbers. Keep both sets when resolving.
• Stacked draft: the retention PR is stacked on this branch. Retarget it to main before merging this one. The repo auto-deletes merged branches, and deleting a base closes the PRs stacked on it.

Follow-ups

• Retention: the stacked draft PR (cap, TOFU and privacy trade-offs for your decision). It's what bounds C, which still drives first-contact writes and launch decode.
• KeychainManager.saveData: still deletes, then adds. A focused SecItemUpdate change would close that window.

🤖 Generated with Claude Code