[rikedyp]

Breaking API changes

• GET /videos - the presenter query parameter is now presenter_id. Searches across multiple presenters return videos by any (OR semantics).
• GET /videos - no longer 404s when nothing matches; it returns an empty
  page. Invalid parameters are 400 (was 422), and page, per_page, sort
  and the date range are now validated rather than silently coerced.
• GET /events - has_videos must be true or false; anything else is a
  400 instead of being treated as "all events".
• GET /videos/{youtube_id}/recommendations - default n is now 5 (was 1).
• Error envelope reshaped error is the HTTP status code as an integer (was
  a machine-readable string). Every error also carries request and
  query_parameters; detail and report_url appear on 500 only, instead of
  being concatenated into message.
• PATCH /crud/{table}/{id} - returns the updated record as a single object,
  not a one-element array. A non-numeric ID is 400; no matching row is 404.
• POST /crud/{table} - an empty JSON array is now 400.
• DELETE /crud/{table}/{id} - 409 removed from the spec.
• Conditional requests - read endpoints send ETag, Last-Modified and
  Cache-Control, and answer 304 to If-None-Match / If-Modified-Since.
  CRUD endpoints send Cache-Control: no-store.
• 503 while the cache is unavailable - read endpoints return a retryable
  503 where they previously failed with a 500.
• Pagination - links report the requested per_page rather than one capped to
  the result count, and an empty result set is page 1 of 1.
• Timestamps are UTC.
• New GET /admin/refresh/status; POST /admin/refresh now answers 202
  with a monitorUrl pointing at it.

Serialised cache refresh

ADMIN.Refresh takes a refresh lock and hands the work to ADMIN.RunRefresh:
one refresh runs at a time with a single pending slot, and a stale lock is
recovered automatically. This removes the races and clobbering that could lock
the system up or let a request read a half-updated cache. A new REFRESH test
suite covers the lock, the status endpoint, and the accepted/pending paths.

Cold start from a persisted cache

The built cache is written to disk (ADMIN.WriteCache / ReadCache, validated by
CACHE.Valid) and re-read at startup, so a fresh process serves requests
immediately instead of failing until the first build completes.

OpenAPI spec conformance

The spec and the implementation were reconciled:

• Corrected parameter and header
  definitions
• Shared conditional-request and cache-header components in
  SPEC.apln
• Richer error envelope
• Proper schema for /teapot.

Other fixes

• Clearer authentication errors
• Static files fall through to the normal 404
  handler
• SQL transactions are committed explicitly and only signal on a failed
  reconnect
• Failed migrations are no longer recorded
• Log output is size-capped.

Performance

• Search indexing now only stems unique corpus terms, reducing external calls.
• Search index lookup function is now explicitly hashed, rather than relying on index∘⍳

Tests, CI and docs

Test framework reworked around assertions that report their failures, suites run
against a clean database, cache builds are exercised in an isolate under test,
and Jenkins runs are timeout-bounded and fail fast on a bad cache build.
Obsolete CI/test-endpoints.sh removed.

[rikedyp]

Closes #134 #137 #157 #159 #160 #161 #162