[efiten]

Closes #764

What changes

createRowAttributor (web/rowattribution.js):

• A slice gets at: askedAt, the moment it was asked for, instead of now() after the answer. A slow answer then ages from the question.
• An answer whose askedAt is older than the current slice does not replace it. That call still attributes its own rows with its own answer, since the newer slice need not cover them (second commit, after review).

Why

The flaky CI test relaynames.spec.js:107 (on PR #762) comes down to this:

• At boot the ticker renders before the role is known.
• applyRole resets the attributors and refetches (web/map.js:610, :679).
• If that registry answer only lands after the test jumps the clock 66 s, the empty slice is stamped after the jump and counts as fresh for another minute.

Tests

• Three unit tests in rowattribution.test.js:
  • a slow answer that lands after the TTL is stale on the next ask;
  • a late older answer does not replace a newer slice;
  • a late older answer still attributes its own rows outside the newer slice.
    All three failed without the fix (called 1 times / rule: estimate).
• A deterministic e2e probe that holds every empty registry answer until after the clock jump: without the fix 1 of 3 runs failed, with the fix 12 of 12 passed. I did not add the probe as a permanent test; the unit tests cover the rule.
• vitest 873/873, eslint clean, relaynames.spec.js 7/7.
• Full Playwright locally: 336 passed, 1 failed (zoomfade.spec.js:90). That test does not touch the registry, and it also failed once in a full run on the branch of fix(web): keep the bar's labels on one line at tablet widths #759 while passing 8 of 8 alone.

Not done

• Left out of scope (pre-existing, from the review): an answer asked before reset() that lands after it can still fill the slice, and a late failed answer sets failedAt without comparing askedAt.
• The original flake rate is low: 2 of 40 locally, then 0 of 120. A rate measurement before and after therefore proves little; the probe is the evidence.
• No changelog entry. The effect for a reader (names up to a minute late after a slow registry answer) occurs rarely and is not a feature.

🤖 Generated with Claude Code

https://claude.ai/code/session_01TDXX3ZV8LEdZgVTNFYFvZi