fix: #172 follow-ups after #169 (analytics warm-up retry, drawer close contrast, IATA oldest test, RX fit race)

[dborup]

Summary

Four follow-ups after #169, as listed in #172. Each item has its own test commit and fix commit, except item 3, which is test only.

Relates to #172
Relates to #150

1. Analytics outlasts the server warm-up (item 5 of #150)

Since #145 the server answers 503 + Retry-After: 5 on /api/analytics/{rf,topology,channels} until its background load is done, for up to the 60 s force-open (cmd/server/analytics_warmup_1659.go). api() gave up after 6 attempts (about 30 s), so the page showed "Failed to load".

• api() takes retry503: false for a caller that retries on its own. Its errors now carry the HTTP status, and a 503's valid Retry-After as retryAfterSeconds. Other callers keep the old 6-attempt loop.
• loadAnalytics() retries a 503 on the server's interval while the next attempt starts within 120 s of the load's start, and shows a "still loading" status (role="status") meanwhile. Only after that does it show the error.
• Like the distance tab (fix(analytics): treat lazy distance-index 202 responses as transient #120), each load and destroy() bump _loadGen and clear the single retry timer. A superseded load or a page the user has left never renders.
• The retry-delay helper is now shared with the distance tab (_retryAfterDelayMs).

The 120 s cap is a constant. Per AGENTS.md rule 8, exposing it in the customizer is left for later.

Perf: this is not a hot path. During a warm-up, each open analytics page sends at most three requests every 5 s, for at most 120 s. Successful endpoints are cached by api() and are not fetched again.

2. Drawer close-button contrast in every preset

The close button used --nav-text-muted on --nav-bg2. That falls below 4.5:1 in forest/light (3.38), forest/dark (4.05), sunset/dark (3.03) and mono/dark (3.67).

The fix gives the button --nav-text, the same colour as the title next to it, which passes on that background in every preset and theme. It is one declaration on one control, so it is the least invasive option. Changing --nav-text-muted in the four presets would also recolour every inactive navbar link and drawer item, and it would not help an operator's own theme. Hover and focus are still visible through the hover background and the focus outline.

3. IATA throttle: oldest refresh after a partial sweep (test only)

TestIATADropThrottleNoResweepAfterPartialSweep does a partial sweep, refills the table, and sends 100 new regions before the next deadline. Exactly one sweep may happen in that window, and exactly one more at the deadline. No production change was needed.

4. RX Coverage: stale fitToObserver response

A fit request now gets a sequence number, like the coverage and leaderboard requests in #169, and it only applies while it is the latest fit (isLatestFit). A newer fit, a change of days and the All button each bump the sequence, so a stale response, or its error path, does nothing.

Tests

Item	Test	Master	Branch
1	test-1659-analytics-warmup.js (#172 part, 5 tests)	5 red	12/12
2	test-nav-drawer-1064-e2e.js step (k), 8 presets × light/dark	4 red (25/29)	29/29
3	TestIATADropThrottleNoResweepAfterPartialSweep	green (code was correct)	green, also with -race
4	test-issue-124-rx-coverage-viewport.js 14a–c	3 red (13/16)	16/16

Mutants (each run in a scratch copy) and the test that catches each one:

• analytics cap set to 30 s → the 45 s warm-up and the permanent-503 tests;
• destroy() no longer cancels → the leaving-the-page test;
• no generation check on success → the slow-superseded-load test;
• the old api() retry loop kept → three tests;
• close button on --text-muted → step (k), 10 combinations red;
• t.oldest = oldest removed → the new IATA test (101 sweeps instead of 1);
• the fit response checks only isLive → 14a–c;
• setDays does not drop the fit → 14a;
• All does not drop the fit → 14b.

The following runs were identical on master and branch:

• test-frontend-helpers.js: 705 passed, plus the 2 known favStar failures.
• test-analytics-channels-integration.js: the same single failure (channels.js sidebar link).
• test-e2e-playwright.js against a local fixture server: stops fail-fast at "Version info lives on Perf dashboard".

scripts/check-xss-sinks.sh --diff origin/master and scripts/check-css-vars.js are clean. The fork guards in deploy.yml are unchanged: 9.

Browser validation (local server, scratch proxies)

• Analytics: a proxy answered 503 + Retry-After: 5 for 45 s. On desktop and mobile the page showed "still loading", then the data at about 45.9 s, and never an error. Master showed "Failed to load" at about 30.8 s.
• Drawer: forest/light, sunset/dark and mono/dark measure 6.39, 9.37 and 17.4:1 for the close button (screenshots taken).
• RX Coverage: a proxy held the observer extent for 4 s. Switching days, choosing All or choosing another observer while that request was in flight never moved the map to the stale extent, on desktop or mobile. On master all six cases moved the map.

🤖 Generated with Claude Code

[dborup-agent]

Review feedback addressed (commits ae27ab35, 45d9e855, fe88e842, d7cd59b2)

1. _inflight ignored retry503 (fe88e842). The in-flight key is now the path plus the flag. A default caller keeps the bare path, so it no longer joins a retry503:false request and gets its 503 after 1 s. A retry503:false caller no longer joins a retrying request and waits ~31 s. Tests: both directions, plus "same flag still shares one request" (test-1659-analytics-warmup.js).
2. Test gaps (ae27ab35, test only; it passes on the existing code and catches the named mutants):
   • a 500 and a 404 show the error at once, with no retry and no pending timer (J19);
   • Retry-After 1 / 12 / 60 / abc → retry after 1 / 12 / 30 / 5 s, and the status text says so (J11b, J4);
   • the loading state is role="status" and not an alert (J8).
3. Tab click during loading threw a TypeError (d7cd59b2). The six tabs that render from the shared load (Overview, RF, Topology, Channels, Hash Stats, Hash Issues) show the load's current status (loading, still loading or the error) until the data is there. The load then renders the selected tab. The load also writes its status only while one of those tabs is shown, so its 503 retries no longer overwrite a tab that fetches its own data (on 1c510d4 the Nodes and Roles tabs got overwritten). Tests: a click on each of the six tabs during a 45 s warm-up must not throw, must show the loading state, and must render that tab after the warm-up. Another test checks that a self-fetching tab is left alone. 45d9e855 gives these tests the real response shapes from the CI fixture server (test-fixtures/analytics-tabs-172.json: arrays cut to two items, names and keys anonymised).

Not changed: the drawer close-button hover colour. Measured on the hover background in 8 presets × light/dark, only --nav-text reaches 4.5:1 everywhere: --accent goes down to 1.14 and --nav-text-muted to 2.73. The hover background change stays the visible difference.

Mutants (each in a scratch copy)

Mutant	Caught by
J13: _inflight keyed on the path alone	the 3 in-flight tests
J19: no e.status === 503 check	"a 500 / a 404 shows the error at once"
J11b: analytics ignores Retry-After (fixed 5 s)	"Retry-After: 1 / 12 / 60"
J4: no err.retryAfterSeconds	"Retry-After: 1 / 12 / 60"
J8: no role="status"	"Retry-After: …" (4) and the 6 tab-click tests
tab click without the guard	the 6 tab-click tests
status written over any tab	"a tab that fetches its own data is not overwritten"

Runs

• test-1659-analytics-warmup.js: 28/28 (was 12). On 1c510d4 the 10 new reproduction tests fail.
• The 85 node test-*.js from the CI unit step: all green, the same as on 1c510d4.
• 34 more app/analytics-related test-*.js: the same results as on 1c510d4. That includes the known failures: test-frontend-helpers.js 705 passed + 2 favStar, and test-analytics-channels-integration.js 1 failure.
• check-xss-sinks.sh --diff origin/master: clean. check-css-vars.js: OK. Fork guards: 9, unchanged.

Browser (headless Chromium, local server, CI fixture; 503s mocked with page.route)

• Warm-up 45 s with Retry-After: 5: the data appears at ~45.1 s with no error, on desktop and mobile. Warm-up 40 s with Retry-After: 45 (clamped to 30 s): the data appears at ~60 s (retries at 30 s and 60 s).
• 20 tab clicks during a 30 s warm-up: 0 pageerrors on desktop and mobile. On 1c510d4 there were 7 (totalTransmissions, snrValues, hopDistribution, channels, distribution). RF, clicked last, renders by itself at ~30.2 s.
• test-nav-drawer-1064-e2e.js: 29/29. test-e2e-playwright.js: 132/135 (3 skipped), the same as on 1c510d4.

🤖 Generated with Claude Code

https://claude.ai/code/session_017vwXk6z7Bcv1AmN2qd6WSZ

[dborup-agent]

Review feedback addressed (commits 837abe98, 7ad6defd)

Test and fixture only. public/ and the other production code are unchanged since d7cd59b2.

1. N1: the self-fetching-tab test proved nothing (7ad6defd). The test wrote its own marker after the click. It now runs for Route Patterns, Roles and Nodes. Each tab gets a small answer from its own endpoint. During the warm-up the test requires that tab's own fetch and its own output. After that it requires that the 503 retries leave the output alone. M8 (LOAD_TABS.has(tab) && removed) now fails all 3 of these tests.
2. N2: no test for the status reset on a new load (7ad6defd). New test: a 500 shows "Failed to load", a region change starts a new load, and an RF tab click during that load must show "Loading analytics", not the old error. RF renders once the new load is done. M6 now fails this test.
3. theme-refresh during a load (7ad6defd). The test environment captures the page's theme-refresh listener on window. Two new tests (Overview and RF) call it during the warm-up. They require no error, an unchanged role="status" loading state, and the data after the warm-up. On 1c510d48 they fail with the reported TypeError (totalTransmissions, snrValues). If the guard is removed, or moved into the click handler so that only the theme-refresh path is unguarded, both tests fail.
4. N3: two real pubkeys in the fixture (837abe98). The two perObserverReach keys are now synthetic 64-hex keys, uppercase like the other observer ids. The values, the key order, the shape and the file size are unchanged. I checked every key and value against e2e-fixture.db: nodes.public_key, observers.id, transmissions.hash / from_pubkey / pubkeys in decoded_json, neighbor_edges.node_a/b, node and observer names, and the 64-hex keys in test-fixtures/*.sql. Real ids: 2 before, 0 after.
5. Extra: the gen check in the error path had no test. New test: a slow 500 from a superseded load must not be shown during the newer load.

Mutants (each in a scratch copy, test-1659-analytics-warmup.js)

Mutant	Caught by
M8: guard without LOAD_TABS.has(tab)	"clicking subpaths / roles / nodes during the warm-up renders that tab…" (3)
M6: no _loadStatusHtml = LOADING_HTML on a new load	"after a failed load, a data tab clicked during the next load shows Loading…"
renderTab guard removed	the 6 data-tab click tests, the 2 theme-refresh tests, the M6 test
guard moved to the click handler (theme-refresh only unguarded)	the 2 theme-refresh tests
_inflight keyed on the path alone	the 3 in-flight tests
no e.status === 503 check	"a 500 / a 404 shows the error at once", the M6 test
load ignores Retry-After (fixed 5 s)	"Retry-After: 1 / 12 / 60"
no err.retryAfterSeconds in app.js	"Retry-After: 1 / 12 / 60"
load status written over any tab	the 3 self-fetching-tab tests
no role="status"	"Retry-After: …" (4), 6 tab-click tests, 2 theme-refresh tests
no gen check after the responses	"a slow response of a superseded load…"
no gen check in the error path	"a slow error of a superseded load…" (new)
warm-up cap 30 s	12 tests, including "a warm-up longer than 30 s…"
destroy() keeps the retry timer	"leaving the page during the warm-up…"

All 14 mutants are caught on the branch and on a local merge with origin/master. Against the old test file, M8, M6, "guard in the click handler" and "no gen check in the error path" survive.

Runs

• test-1659-analytics-warmup.js: 34/34 (was 28).
• On the branch, test-app-api-inflight-cleanup-rejection.js 8/8, test-warmup-banner.js 13/13 and test-issue-1375-scope-stats-fetch.js 9/9 pass. test-frontend-helpers.js fails 2 (favStar), and test-analytics-channels-integration.js fails 1, both already fixed on master.
• On a local merge with origin/master (cbd562cd): all of the above are green, including test-frontend-helpers.js 707/707 and test-analytics-channels-integration.js 24/24.
• Local full validation at 7ad6defd (Go with -race, the JS unit tests in the CI list, Playwright E2E on the CI fixture): 204 PASS, 0 FAIL.
• Fork guards: 9 github.repository == 'Kpa-clawbot/CoreScope', unchanged.

🤖 Generated with Claude Code

https://claude.ai/code/session_01TdKwzUPntKQHjQEKpMXefc

[dborup]

Review — CS-Cloud re-review PR#175 runde 3 — head 7ad6def

Dom: APPROVE with nits

Delta review of d7cd59b2..7ad6defd (2 commits, 2 files, +113/−15). Read-only: I read the code through git archive copies and ran everything in a scratch directory. The head was 7ad6defd07749c00d943226eea138b902b25df30 before and after the review (git ls-remote). Tags: [T] test, [A] analysis, [K] not checked.

Checks

Check	Result
No changes in public/	Yes [T]. git diff --name-only d7cd59b2 7ad6defd lists no public/ file, and the public/ trees of both commits are identical. The round only changes test-1659-analytics-warmup.js and test-fixtures/analytics-tabs-172.json.
Fixture shape unchanged	Yes [T]. The only change is the two topoData.perObserverReach keys. The shape is identical with those two keys compared by position, and no leaf value changed.
Fixture free of e2e-fixture.db pubkeys	Yes [T]. The fixture has 22 distinct 64-hex strings. I extracted every 64-hex string from every text column of test-fixtures/e2e-fixture.db (258) and compared case-insensitively: 0 in common, as keys or as values. Searching the raw bytes of the DB file also finds none. The same scan on round 2 found 2 in common (the two old keys), so the fix did what it should. The DB file is unchanged.
M8 caught	Yes [T]. See the table below.
M6 caught	Yes [T].
Guard in the theme-refresh path caught	Yes [T]. See the table below.
CI on 7ad6def	Green [T]. I read the check runs through the GitHub API; the API reported 7ad6defd as the PR head. Go Build & Test: success. Playwright E2E Tests: success. Build & Publish Docker Image: success. Staging deploy, badges and release artifacts are skipped, as expected for a draft PR.

Test results [T]

• test-1659-analytics-warmup.js: 34 passed, 0 failed (round 2: 28). I ran it 5 times with the same result; it uses a fake clock and a fake Date.
• test-packet-filter.js 92/0, test-aging.js 19/0, test-app-api-inflight-cleanup-rejection.js PASS, test-warmup-banner.js 13/0, test-issue-124-rx-coverage-viewport.js 16/0, test-issue-120-distance-building.js 13/0.
• test-frontend-helpers.js: 705 passed, 2 failed. These are the known favStar failures, identical on master.

Mutants [T]

Each mutant was applied to a scratch copy of the head and run against both the round-2 test and the round-3 test. "Survived" means the test file stayed green.

Mutant	Round-2 test	Round-3 test
M8a: the renderTab load-status guard applies to every tab, so own-data tabs show the status	survived	killed by clicking "subpaths"/"roles"/"nodes" during the warm-up renders that tab, and the retries leave it alone
M8b: the load status is written over a tab with its own data	killed	killed (same 3 tests)
M8c: subpaths added to LOAD_TABS	killed	killed (subpaths test)
M6: a new load does not reset the status to "Loading"	survived	killed by after a failed load, a data tab clicked during the next load shows "Loading", not the old error
T1: the renderTab guard is removed	killed (6)	killed (10), including both theme-refresh tests
T5: guard bypassed only on the theme-refresh path (the handler renders overview and rf without the guard)	survived	killed by theme-refresh on "overview"/"rf" during the warm-up keeps the loading state, then the data renders
T3: the theme-refresh listener is not registered	survived	killed (both theme-refresh tests)
T2: guard only for overview	killed	killed
T4: collisions missing from LOAD_TABS	killed	killed

So M8a, M6, T3 and T5 survived round 2 and are now killed. T1 and T2 were already killed by the tab-click tests, because they share the renderTab guard with the theme-refresh handler. T5 shows that the new theme-refresh tests catch a guard bypass that exists only on the theme path.

Nits

1. [K] The fixture still holds 20 other 64-hex values (hop pubkeys, observer_ids and observers[].id). None is in e2e-fixture.db or anywhere else in the repository, which is what this round asked for. I cannot tell from the repository whether they came from a real network capture. If "fully synthetic" is meant strictly, they could be replaced the same way as the two keys. This is not blocking.
2. [A] The round-3 tests add OWN_TABS and pageEnv options (realData, status, retryAfter) inside one 700-line test file. That is fine now. If more tabs are added later, moving the fake DOM and clock helpers to a shared test helper would help.

Not checked

• [K] The browser behaviour of the round-3 tests. They run in a vm sandbox with a fake DOM, so they prove the logic and not the rendering. My earlier browser checks covered 1c510d48, not this head.
• [K] The Go, CSS and rx-coverage changes of earlier rounds were not re-run on this head. None of those files changed in this round.

---

Generated by Claude Code