Raise testem's disconnect timeout; trim DEFECTS.md to what is open

[roncodes]

Two things: a CI-breaking regression from my own earlier fix, and the DEFECTS.md cleanup.

1. The run exits 1 even though every test passes

The first full run after #164 and #165 merged:

# tests 5197
# pass  5196
# fail  1

not ok 5197 Chrome - error
  Error: Browser timeout exceeded: 10s

All 5196 real tests passed and the coverage report was written. The run failed anyway.

This is a regression from #165's fix. Testem.afterTests makes testem wait for the coverage POST — that is the entire point of it, and it is what stopped the uploads being truncated. But testem's browser_disconnect_timeout defaults to 10 seconds, and the payload is several megabytes once forceModulesToBeLoaded() has run. My verification run at 5130 tests was under the threshold; #164's added code pushed it over.

It would have failed CI on every run, with a message that names neither coverage nor the upload.

Fix: browser_disconnect_timeout: 120 in testem.js, alongside the existing browser_start_timeout: 120. That timeout exists to catch a hung browser; waiting on a deliberate, bounded upload is not that.

2. DEFECTS.md: 746 lines → 140

Seventeen of nineteen entries were fixed, and fixed entries are noise in a worklist. Removed, along with the appendix mapping an old numbering scheme nothing uses any more and the kickoff brief for #15 — which #164 has now implemented.

The full file is preserved at c9bc486 and linked from the header, so anything in the repo citing an old defect number still resolves. I verified the link actually returns the file rather than trusting the path — the first version pointed at the monorepo path and 404'd, since the remote is the standalone repo.

What is left:

• One open entry — small styling updates for tag input component, and added param to adj… #18, the ±1 branch wobble between identical runs.
• The settled list — behaviour that is deliberate and has been "fixed" wrongly before: sticky columns, the resource panel's unwired save task, the removed panel callbacks, custom-fields staleness.
• Reference material that is still forward-looking — why the remaining coverage gaps are where they are, and the habits worth keeping.

One thing added rather than removed

A short section on the trade coverage collection is balanced on, because the two entries it replaces were the same problem pulling in opposite directions:

• Post from QUnit.done → nothing waits → the upload truncates and no report is written, silently (upstream #420).
• Post from Testem.afterTests → testem waits → its disconnect timeout must outlast the upload, or you get this PR's failure.

The payload size is the real variable, and forceModulesToBeLoaded(filterFunction) is the lever nobody has pulled: we currently force-load and transmit coverage for workspace siblings like @fleetbase/ember-core that scripts/check-coverage.js then discards.

Verification

Full run before this change: 5196 pass, 1 error, exit 1. A confirming run with the timeout raised is in progress and I will post its result here.