Fix what a retroactive review found in what the site publishes

[baz8080]

Slice 3 of the retroactive review, following #48, #51 and #52: a whole-file review of esb_site/render.py, site.html, the three static templates, site.css and __main__.py, then three reviews of this PR. Each commit holds one fix, and each fix's test was checked to fail without it.

The fixes

1. Events over before the first poll were counted outside the county page. Area pages, search, repeat chains and the build summary still saw them: 14 outages, a page for an area whose only event was one of them, and a repeat fault reading "outage 2 of 2" with no outage 1. They are now dropped once, in model.load_outages, by the end the merge settles on and before the chains are labelled.
2. CSV formula injection. A text cell whose first non-space character is =, +, - or @, or that opens with a tab or carriage return, now gets a leading apostrophe.
3. --now dropped an offset instead of converting it.
4. The county row spliced the name into an inline onclick JS string. It now uses data-county and one delegated listener. The area page's link into the app encodes the county.
5. A malformed share link (#county/%E0) blanked the app. It now falls back to the overview.
6. "1 customers" is now singular in the timeline and the fault-spots card.
7. The nearby-area links now use area_path.
8. A month the data hasn't reached says "no data yet" (owner's call), never zeros.
   • model.month_watched decides it once, and the app gets it as D.nodata.
   • The county table shows the caveat, a grade chip with its reason, and dashes. The gate sentence says "There is no data yet for October 2026", and no grade date ships for that month.
   • The app shows it in the headline, the tiles, the county rows and the county view.
   • model.overlaps rejects an empty window, so nothing is counted or filed in such a month.
   • "So far" belongs only to the newest listed month, and only once data reaches it.
   • Recorded in notes/design-alignment.md, with a CLAUDE.md Settled row.

JS was checked in Chromium with a normal build and one dated 00:20 on 1 October:

• the headline and basis wording is right for both months;
• row clicks work and the bad link recovers;
• there are no page errors;
• all 856 Dublin and Cork records render identically from JS and Python.

CI runs no JS, so the call sites are held by text tests.

Reviews of this PR

• First review, 9 findings, all fixed. Chiefly, the first version of fix 1 filtered after the repeat chains were labelled.
• Second review, 10 findings. Chiefly, "no data yet" used different rules on the page and in the app. It now uses one rule, shipped to the app.
• Third review, of that fix alone, 8 findings, all fixed.
  • The real bug: the one-microsecond observed_month change left "so far" keyed on it, so at a month start the app called September "so far" while the page had closed it.
  • A structural fix: the empty-window guard was copied into three places; it's now model.overlaps.
  • Smaller fixes:
    • a grade date shipped for a no-data month;
    • the table row was untested and had an unstyled grade cell;
    • an old cached data.js would bring the zeros back;
    • two comments restated the settled decision.

Not taken:

• A whole month the collector missed reads as watched once data resumes after it. That's the in-window gap notes/grading.md names.
• The merge filter dropping an event whose lingering sibling ended after the first poll. That's the merge's own rule.
• Folding the inline plural suffixes into one helper: wider than this PR.

Effect on the site

Every published county and national figure is identical to main. The area directory, area pages and search lose the 14 events over before collection began.

Checks

• ruff check: clean.
• unittest discover with ESB_DATA_DIR set: 346 tests OK.

🤖 Generated with Claude Code

https://claude.ai/code/session_01UqWuH6frirpbQDKnF5rEdT