docs: fix the last drifted claims, and make usage-line drift impossible

[FZ2000]

The audit's remaining findings, plus a structural fix so this class stops
recurring.

CONSOLE TRANSCRIPTS THE CODE CANNOT PRODUCE

docs/commands/README.md promises "Output shown in console blocks is
real output from the command, not an illustration." Three blocks were
not:

• getting-started.md showed feed sync stopped: complete. sync.py
  rewrites a surviving "complete" to TIME_BUDGET_EXHAUSTED before the
  summary prints, precisely so a truncated run cannot look finished —
  so that string never reaches a terminal. Now feed exhausted.
• troubleshooting.md showed a 503 as [error] HTTP 503 from <url>: Service Unavailable followed by re-run with -v for the full traceback. Both lines are wrong: _advice_for_http prints
  "DeviantArt is having trouble (HTTP 503 Service Unavailable)", and
  for a 5xx offer_traceback is False so the second line never
  appears at all — and where it does appear it says "re-run with -v to
  see the request that failed". Replaced with the real two lines,
  captured by stubbing a 503.
• auth.md showed [error] refresh failed: HTTP 401 .... The string
  "refresh failed" does not exist anywhere in dacli/. The real message
  names the refresh token and tells you to run da auth.

USAGE LINES — and the reason they drifted unnoticed

Three were stale: da daily had gained --json, da search user had
gained --json, and da sync watched had gained a mutually exclusive
group, so argparse prints [--user USER | --via-feed] where the doc
still showed [--user USER] [--via-feed].

tools/check_doc_flags.py already walks the parser, but it only ever
compared the flag tables. The usage: line above each table — the part
a reader copies verbatim — was checked by nothing. It is now: every
usage: da ... block in docs/ is resolved to its subparser and compared
against format_usage(), whitespace-normalised and first line only,
since argparse wraps to terminal width. 31 usage lines now verified
alongside the 67 flag rows.

The check earned its place immediately: I had fixed two usage lines by
hand from the audit's list, and it found a third I had missed
(da search user). Negative control confirmed — reverting one drifted
line takes the count from 0 problems to 1.

Two bugs in my own addition, both caught before pushing: I inserted the
call after the loop that prints problems, so usage findings were
counted but never displayed; and the summary line still said
"flag-table problem(s)" for what are now documentation problems
generally.

ADR 0007 had drifted past its own refactor: "thirteen command handlers"
(that is the top-level subcommand count; there are 29 handlers) and "the
largest module is 848 lines" (sync.py is ~1,200, and it is the walk plus
its checkpointing — not usefully divisible, which is worth saying rather
than quoting a number that will drift again).

Verified: ruff, mypy, doc references, doc flags + usage lines, codespell
all clean; lychee 194 OK / 0 errors with fragments; 868 tests pass.