fix(app): a requested refresh can coalesce onto an older in-flight api() request (invalidateApiCache does not clear _inflight)

[dborup]

Relates to #232, #238

Problem

invalidateApiCache('/channels') in public/app.js clears only the TTL cache. It does not clear api()'s _inflight map.

If a /channels request is already in flight when a refresh is requested, the refresh coalesces onto the older request and can render without the newly approved channel until the next refresh. A refresh is requested by onApproved after an admin decision or an auto-approve (#238). A request can already be in flight because of a region change, the encrypted toggle, or a second approval within one round-trip.

This existed on the admin path before #238. #238 adds a second trigger, and the window is small: a status poll waits at least one poll delay. It was found in the #238 review and has not been reproduced.

The same review found a test gap: nothing pins that ChannelProposals.unmount() cancels the suggest poller. A mutant that drops state.suggestPoller.cancel() survives the proposals and client-state tests.

Proposed fix

• Give api() a way to bypass _inflight for an explicit refresh, e.g. api(path, { bust: true }), or an in-flight generation that a refresh bumps. Use it from loadChannels(true).
• Keep the request token from fix(channels): drop stale channel-list responses and show the unread badge on mobile (#154, #155) #225 (channelsRequestId), so that an older response is still dropped.
• Add a unit test with controlled api() timing: a refresh during an in-flight request must render the newer data.
• Add a test that unmount() cancels the suggest poller.

Acceptance

• A refresh requested while a /channels request is in flight results in a list that includes the newly approved channel, with no extra request in the normal case.
• Both new tests fail before the fix, and the mutants named above are killed.

[dborup-agent]

Plan — CS-pve-agent2 #243

Branch codex/issue-243-api-inflight-refresh from origin/master 0572e7f9. Draft PR only.

Fix

public/app.js api(): bust: true will also skip _inflight, not just the TTL cache.

• A bust call always starts its own fetch and takes the in-flight slot for its key. Callers that come later and do not bust join the newer request, not the older one.
• Cleanup and cache writes only apply to the request that still owns the slot:
  • .finally() deletes the _inflight entry only if it is still this request's own promise. An older request that settles late must not remove the newer entry.
  • An older, superseded request no longer writes _apiCache. Otherwise its late response would overwrite the newer data for the rest of the TTL.

public/channels.js:

• loadChannels(silent, opts) passes opts.bust through to api('/channels…', { ttl, bust }).
• onApproved calls loadChannels(true, { bust: true }). This is a one-line change inside the onApproved block.
• invalidateApiCache('/channels') stays, because it also clears the cached lists for other region and includeEncrypted variants.
• The channelsRequestId token (fix(channels): drop stale channel-list responses and show the unread badge on mobile (#154, #155) #225) and latestChannelsLoad are unchanged, so an older response is still dropped without a render or a reconcile.

Effect on other api() callers

• Callers that do not pass bust behave as before, byte for byte: same dedup, same TTL hits, same cleanup. In their case the in-flight slot is always their own promise, so the new ownership checks are always true.
• The only other bust caller is the manual refresh in observers.js (loadObservers({ bust: true })).
  • Today it joins an in-flight auto-refresh. With the fix it fetches fresh data, which is what its fix(observers): show "Last updated" timestamp on aggregate header (closes #1562) Kpa-clawbot/CoreScope#1563 request-id guard already expects.
  • That is the one behaviour change, and it only happens when a manual refresh overlaps an in-flight /observers request.
• The other channel triggers do not bust and are unchanged: region change, the encrypted toggle and init().
• Normal case: an approval with nothing in flight still makes exactly one /channels request.

Tests (written first, red on master)

1. New test-app-api-bust-inflight-243.js, which loads the real app.js with a stubbed fetch:
   • a bust during an in-flight request fetches again and gets the newer data;
   • a later non-bust call joins the newer request;
   • the older request settling late neither removes the newer in-flight entry nor overwrites the TTL cache;
   • with nothing in flight a bust makes one fetch;
   • non-bust dedup and TTL behaviour are unchanged.
2. test-channels-client-state-152.js, using the real channels.js with the real app.js api() and controlled fetch timing:
   • a /channels request is in flight when onApproved fires, and the rendered list contains the approved channel in both resolve orders;
   • the older response is dropped (fix(channels): drop stale channel-list responses and show the unread badge on mobile (#154, #155) #225);
   • in the normal case there is exactly one request.
3. test-channel-proposals.js: ChannelProposals.unmount() cancels the suggest poller. No timer stays pending, and no status request is made after the unmount.

Mutants:

• bust ignored for _inflight in api();
• onApproved without bust;
• state.suggestPoller.cancel() removed;
• the ownership checks removed.

Validation

• sh test-all.sh, node test-frontend-helpers.js, node test-channel-proposals.js, node test-channels-client-state-152.js, and the observers tests.
• test-channel-proposals-e2e.js and the channel E2Es against a local Go server on e2e-fixture.db, prepared as in CI.
• scripts/check-xss-sinks.sh --diff origin/master. The fork guards are unchanged.
• No CSS, colours or new requests per item. Not a hot path: one map lookup per api() call.
• Config/customizer: no new configurable values.

Overlap with #257

#257 replaces the onApproved block with refreshChannelList(). Whichever PR merges second needs a one-line resolution: loadChannels(true, { bust: true }) goes into refreshChannelList. New tests are placed away from #257's test hunks.

[dborup]

Fixed by #263 (merged as f91339f2).