Atmosphere triptych — 3 curated picks with variety engine (#76)

* feat: atmosphere triptych — 3 curated picks with variety engine

Replaces single-recommendation Atmosphere modal with a 3-card triptych:

Backend:
- New ATMOSPHERE_PROMPT requesting 3 ranked picks with lateral-thinking guidance
- Recommendation history file (atmosphere_history.json) tracking picks over time
- Inverse-frequency weighted pre-filter samples ~20 candidates, avoiding repeats
- random.shuffle() eliminates insertion-order bias in LLM context
- temperature 0.9 and max_tokens 600 for atmosphere calls only
- Fallback padding ensures 3 recommendations even if LLM returns fewer

Frontend:
- 3-card triptych layout (820px max-width, responsive stack at <=600px)
- First card marked Curators pick with accent top-border
- Staggered reveal animation (150ms intervals)
- Each card: thumbnail, wiki-linked title, curator note, vibe pills, Display button
- Explored N of M counter below cards
- atmosphereRetry() excludes all 3 picks at once
- localStorage persistence with 24h TTL for exclusion list
- closeAtmosphere() no longer clears exclusions

Tests:
- Updated test_full_pipeline for 3-recommendation response schema
- New test_fallback_padding and test_history_persistence
- Updated XSS test for safeTitle variable in pick cards

Closes #65

* fix: deduplicate repeated content_ids in atmosphere recommendations

Track used_ids during LLM result parsing to skip duplicate content_ids
before the padding step. Prevents duplicate cards when the model returns
the same artwork more than once.

Adds test_dedup_repeated_ids to verify 3 unique picks are returned even
when the LLM repeats an ID.

Merge pull request #75 from danmunz/fix/prefetch-timeout

Merge pull request #74 from danmunz/fix/thumbnail-resilience

fix: thumbnail resilience for large collections (#67)

Merge pull request #72 from danmunz/fix/thumbnail-prefetch-throttle

fix: throttle background prefetch and add circuit breaker (#67)

feat: guided Vision API setup with key validation (#71)

- Add VisionApiError with structured error parsing (extracts Google's
  project-specific enable URL from 403 responses)
- Add POST /api/ai/test-vision endpoint for key validation
- Replace 'Get a key' hint with numbered setup guide (Enable API →
  Create key → Paste) and inline Test button with ✓/✗ feedback
- Propagate vision_error in analysis metadata; show warning in UI
  when Vision silently fails
- Auto-test after saving a new key; support test-before-save flow
- 7 new tests for test-vision endpoint

fix: non-blocking thumbnail prefetch and WebSocket staleness recovery (…

…#67)

* test: add diagnostic and reproduction tests for issue #67 thumbnail failures

Add 20 tests across two files that identify and reproduce the root cause
of thumbnail loading failures on large catalogs (500+ artworks).

Root cause: client-server timeout race condition. The server-side
individual fallback (PR #69) takes 53s+ for a batch of 20 IDs when the
TV is struggling, but the frontend's AbortController timeout is 30s.
The client always aborts before receiving the response, never sees the
fallback:true flag, so the progress bar never appears. Three retries
create a ~96s loop ending in 'Tap to retry' on all thumbnails.

test_issue67_diagnosis.py (14 tests):
- Thumbnail key matching with samsungtvws format
- Individual fallback cascade timing (20 TV calls per failed batch)
- TV operation timeout and D2D socket orphaning
- Lock starvation during concurrent fallback operations
- Large catalog scenario (63 TV hits per batch of 20)
- Client timeout race: proves 1 failing ID = 58s worst-case (>30s)

test_reproduce_nitrowolf.py (6 tests):
- Server response exceeds client timeout (53s vs 30s) — matches logs
- Click-to-retry reads from disk cache (2.6ms) — matches user report
- Progress bar never shown (client aborts before response)
- Three-attempt retry loop (96s wall time → 'Tap to retry')
- Concurrent batches compounding the timeout problem
- Settings endpoint blocked by TV lock during fallback

Uses 50x time scaling for fast test execution while preserving the
timing relationships that cause the failures.

Refs: #67, #68, #69

* fix: address Codex review feedback on PR #70

- Mark reproduction test classes with @pytest.mark.xfail(strict=False)
  so a production fix for #67 won't break CI. Tests currently show as
  XPASS (bug still exists); once fixed they'll become XFAIL (expected).

- Replace time.sleep(10) in hanging_fn with threading.Event.wait(2)
  and cancel.set() after assertions, so the executor thread exits
  promptly instead of stalling the suite for 10s.

* fix: non-blocking background prefetch for thumbnail fallback (#67)

Replace synchronous individual thumbnail fallback (which blocked 53s+
for large catalogs) with asyncio.create_task() background prefetch.

Server changes:
- Add _prefetch_thumbnails() background task that fetches thumbnails
  individually via _tv_op() and caches to disk
- Add _thumb_prefetch_in_progress dedup set to avoid redundant fetches
- get_thumbnails_batch() now returns immediately with fallback=True
  and all uncached IDs as missing, spawning background prefetch

Client changes:
- In fallback mode, retry up to 8 times at 3s intervals (vs 3 retries
  with escalating backoff) to pick up newly cached thumbnails

Test updates:
- conftest: reset _thumb_prefetch_in_progress and _tv_lock per test
- test_api_endpoints: verify immediate response + background cache
- test_issue67_diagnosis: update 5 diagnostic tests for new async
  behavior (fast response, background individual calls, disk cache)

* fix: proactively reconnect stale TV WebSocket connections

Samsung Frame WebSocket connections go stale after ~30-60s of
inactivity, causing BrokenPipeError when the user interacts with
the TV after browsing Atmosphere results or other idle periods.

- Track _tv_last_used timestamp, updated after each successful _tv_op
- _ensure_tv_connection checks idle time against TV_CONN_MAX_IDLE (30s)
- Stale connections are proactively closed and reopened instead of
  waiting for BrokenPipeError on the first attempt
- Reset _tv_last_used in test fixtures for proper isolation

* refactor: address PR review comments from Copilot and Codex

- Fix misleading log: 'falling back to individual' → 'scheduling
  background prefetch' (Copilot on server.py:524)
- Fix misleading comment: 'shorter delays' → 'fixed 3s interval'
  with note about normal escalating backoff (Copilot on index.html:2879)
- Replace fixed sleep(0.2) with bounded polling loop in fallback
  tests to avoid CI flakiness (Copilot on test_api_endpoints.py:395,424)
- Rename test_526 → test_524 to match docstring catalog size
  (Copilot on test_issue67_diagnosis.py:281)

Merge pull request #69 from danmunz/copilot/diagnose-issue-67-fix

fix: consistent API shape & Set-based progress tracking for thumbnail fallback

feat: persistent TV connection, /health endpoint, corruption recovery

Closes #67, #57, #63, #19, #47

- Persistent WebSocket connection (11 connections → 2)
- Removed concurrent socket race (_fetch_thumbnails_sync)
- /health endpoint with Docker HEALTHCHECK (503 on errors)
- JSON corruption recovery with _safe_load_json
- Jump-to-current-artwork FAB
- AI settings inline help text

chore: bump version to 1.0.1

chore: bump version to 1.0.0