fix: refresh SCShareableContent on topology change, not per call

[divanshu-go]

Fixes #1.

Problem

SCShareableContent.current forces replayd to enumerate every window, app, and display system-wide, building an NSSet of NSDictionary metadata with O(n²) equality checks. Apple designed it for one-time initialization — but this crate called it on every Window::all() / Monitor::all() / one-shot capture / stream-filter build. For an always-on consumer that's an enumeration every few seconds: replayd pinned at 100% in -[NSSetM addObject:] and jetsam-killed in a loop (#1's profiling), plus the slow replayd memory bloat documented in alt-tab-macos #4194.

A fixed TTL is not the fix — it still enumerates ~17k times/day forever. The prior art in Hyperion #1921 is explicit (m13v's comment): fetch once, refresh only when the window/display topology actually changes.

Fix — topology-validated cache

The snapshot is cached and revalidated per call against a cheap fingerprint from CGWindowListCopyWindowInfo — the pre-ScreenCaptureKit WindowServer query (sub-millisecond, no replayd involvement; the pre-SCK capture backend used it wholesale with zero replayd cost, per #1's own context). The fingerprint hashes every onscreen window's id, owner pid, and title (title changes can flip exclusion-pattern matches downstream), plus the online display count.

The fingerprint is deliberately order-insensitive (per-window hashes, sorted, then combined): CGWindowList returns windows in z-order, and a mere focus switch reorders the list without changing the window set — hashing in list order would turn every cmd-tab into a replayd enumeration, exactly the load this cache exists to remove. Nothing downstream consumes z-order from the snapshot; window focus is derived live from NSWorkspace at Window::all() time.

get_shareable_content()          ── sync; keeps the post-wake 0-display retry loop
get_shareable_content_cached_async()  ── async twin for one-shot capture paths
        └── shared cache: { fetched_at, topology_signature, content }
            hit ⇔ signature unchanged (and < 300s safety-net TTL)

Properties:

• Static desktop → zero replayd enumerations. Not 12/min — zero.
• Focus switches between existing windows → zero enumerations (verified: cached call 2ms after an osascript app activation vs 20–75ms for a real fetch).
• New / closed / retitled window → picked up on the very next call. Fresher than any TTL; exclusion filters see topology changes immediately.
• 300s safety-net TTL bounds staleness the fingerprint can't see; a 5s blind TTL applies only if CGWindowList is ever unavailable.
• 0-display snapshots (SCK's stale post-wake state) are never cached, so the wake-retry logic can't get pinned behind the cache.
• stop_all_streams() drops the cache (wake/display teardown), and hosts can call the exported invalidate_shareable_content_cache() directly.
• Ids and exclusion filters derive from the same snapshot, so they stay mutually consistent. Enumeration semantics otherwise unchanged (full window list, as before).

Measurements (M-series, ~150 onscreen windows)

tests/cache_timing.rs — 20 back-to-back Window::all() calls:

	first call	calls 2–20 (avg)
before (every call = replayd enumeration)	~75ms	~18–75ms each
after (signature-validated cache)	~75ms	0.35ms each (CGWindowList check, no replayd)

Live screenpipe debug engine, 60s run with an ignore pattern configured (RUST_LOG=sck_rs=debug):

fetches=8 frames=7 errors=0
15:33:22.36                                  ← startup
   (43 seconds of static desktop — ZERO enumerations)
15:34:05.85 … 15:34:14.46                    ← real window activity burst

Each fetch corresponds to an actual topology change; idle stretches and focus switches perform none. Streams, exclusion filters, captures, and DB writes fully functional; cargo test suite green.

Honest caveat: on this machine (fast Apple Silicon) replayd never showed the catastrophic spin even uncached — that needs the bigger NSSet from many Electron apps on macOS 14.x. The fix removes the anti-pattern at the source either way, following the issue's prior art.

🤖 Generated with Claude Code