@camstack/addon-export-alexa 1.2.18 → 1.2.19

package/dist/export-alexa.addon.js

@@ -23884,13 +23884,24 @@ method(object({
 /** Playback-speed multiplier for the render (1 = realtime). */
 var ExportSpeedSchema = number().min(.25).max(32);
 /**
-* One dense interval, in SECONDS FROM THE EXPORT'S OWN `fromMs`.
+* One dense interval, in WALL-CLOCK SECONDS FROM THE EXPORT'S OWN `fromMs`.
 *
-* Relative and not absolute epoch on purpose: the renderer's frame-select
-* expression sees ffmpeg's `t`, which starts at 0 for the export's source
-* playlist. Handing it absolute epochs would make every call site responsible
-* for the same subtraction, and the one that forgot would emit a filter that
-* selects nothing — silently, as a uniform timelapse.
+* **Wall clock, not ffmpeg's `t`** — and the recorder translates. A caller
+* derives these bounds from things that happened at a TIME (a track's
+* `firstSeen`), while `t` runs over the source playlist: the concatenation of
+* every segment present for the range, with each recording GAP removed. The
+* two agree only on a window that recorded without one interruption, and only
+* the render side knows the segments, so the translation lives there
+* (`export-dense-map.ts`, addon-pipeline).
+*
+* It was not always so. These seconds were fed to `between(t,…)` verbatim, and
+* on a 10 h window holding 29,393 s of footage every range landed late by the
+* gap accumulated before it — up to 6,607 s, well past EOF. Nothing matched,
+* the video was a uniform timelapse, and the log line reported the five ranges
+* that had been ASKED for (2026-08-13, export `57d14363`, camera 615).
+*
+* Relative and not absolute epoch, because an absolute epoch would make every
+* call site responsible for the same subtraction.
 */
 var ExportDenseRangeSchema = object({
 	fromSec: number().nonnegative(),
@@ -31020,25 +31031,32 @@ object({
 var NativeLeaseAdmissionSchema = _enum(["all", "inferred"]);
 object({
 	/**
-	* How long a retained native frame is served before it counts as a miss.
+	* How many delivered frames the worker HOLDS at once, waiting for each one's
+	* detection result.
 	*
-	* Must cover the FULL late-crop horizon: detection inference + the
-	* cross-process inference-result hop to hub post-analysis + tracking + the
-	* tRPC crop round-trip back. Below ~500 ms the busiest cameras' subject crops
-	* outrun it and fall back to the ≤640 detection frame; above ~3 s the resident
-	* RAM per busy camera grows linearly with no measured hit-rate gain.
+	* This replaced a TTL on 2026-08-13, and the replacement is the whole point:
+	* a time window was never related to the event the pixels were waiting for.
+	* A held frame now lives from delivery until the runner has its `FrameResult`
+	* — at which moment the runner cuts the subject tiles it actually wanted and
+	* releases the frame. The bound exists only so a runner that stops answering
+	* cannot pin RAM: above it the OLDEST held frame is dropped and counted.
+	*
+	* Sizing: the steady state is `inferenceLatency × deliveredFps`, measured at
+	* 40-160 ms × ≤25 fps = 1-4 frames. The default leaves headroom for a hiccup
+	* without ever approaching the old resident set (43 frames × 24.9 MB at 4K).
+	* Raising it does not buy hit rate — it buys tolerance for a slow runner, and
+	* `holdOverflow` on the metrics line is what says you need it.
 	*/
-	ttlMs: number().int().min(250).max(1e4),
+	holdFrames: number().int().min(1).max(64),
 	/**
 	* Hard per-decode-worker RAM ceiling for retained native frames, in MB.
 	*
-	* Intended as a SAFETY ceiling with the TTL as the effective cap — but check
-	* which one is actually binding before reasoning from that. At the shipped
-	* 1024 MB and a 2 800 ms TTL, a 4K camera hits the CEILING first (~43 frames
-	* at ~24 MB each) and the TTL never gets to expire anything; `leaseMb` /
-	* `leaseFrames` on the metrics line say which. When the ceiling binds, a
-	* change that admits fewer frames buys retention WINDOW at constant RAM
-	* rather than giving RAM back — lower this knob if RAM is what you wanted.
+	* Since 2026-08-13 this is a SAFETY ceiling and nothing else: `holdFrames`
+	* is what decides how much is held, and the ceiling is the number above which
+	* something is wrong. Before that it was the effective cap — at 1024 MB with
+	* a 2 800 ms TTL a 4K camera sat pinned at `leaseMb:1020, leaseFrames:43`
+	* with the TTL expiring nothing, which is exactly the confusion the hold
+	* removes. `leaseMb` / `leaseFrames` still say what is resident.
 	* `0` DISABLES the lease entirely and falls the worker back to the tiny
 	* leak-prone GPU surface ring (~85% crop miss; that is what the lease exists
 	* to replace).
@@ -31064,22 +31082,45 @@ object({
 	* there is the signal that some caller names frames outside the inference set
 	* and that this must go back to `all`.
 	*/
-	admission: NativeLeaseAdmissionSchema
+	admission: NativeLeaseAdmissionSchema,
+	/**
+	* RAM ceiling per decode worker, in MB, for the SUBJECT TILES — the
+	* compressed native crops the worker cuts at the moment a frame's detection
+	* result arrives, and keeps long after the frame itself is freed.
+	*
+	* This is the knob that replaced the old retention window, and it buys about
+	* three orders of magnitude more of it: a tile is one subject at native
+	* resolution, JPEG-encoded (~60-120 KB on a 4K person), against ~24.9 MB for
+	* the frame it was cut from. A frame on which nothing was detected costs
+	* nothing at all, which is the real change — the old lease paid per FRAME and
+	* was interrogated per SUBJECT.
+	*
+	* `0` DISABLES tiles, leaving only the hold window and the ≤640 RAM
+	* fallback — i.e. the pre-2026-08-13 miss profile. Set it there only to
+	* reproduce that.
+	*/
+	tileBudgetMb: number().int().min(0).max(1024)
 });
 /**
-* The values in force when the operator has set nothing — byte-for-byte the
-* constants the decode worker shipped with as env-var defaults, so making these
-* settings changed no behaviour on the day it landed.
+* The values in force when the operator has set nothing.
+*
+* `budgetMb` stays at 1024 on the day the hold landed, deliberately: it stopped
+* being the retention window and became the OOM ceiling, and lowering a ceiling
+* in the same change that redefines it would make a regression and a retune
+* indistinguishable. Cut it once `tileHits` / `holdOverflow` have been read on
+* live traffic.
 */
 var DEFAULT_NATIVE_LEASE_SETTINGS = {
-	ttlMs: 1200,
+	holdFrames: 8,
 	budgetMb: 1024,
 	activityMs: 15e3,
+	tileBudgetMb: 64,
 	admission: "inferred"
 };
-DEFAULT_NATIVE_LEASE_SETTINGS.ttlMs;
+DEFAULT_NATIVE_LEASE_SETTINGS.holdFrames;
 DEFAULT_NATIVE_LEASE_SETTINGS.budgetMb;
 DEFAULT_NATIVE_LEASE_SETTINGS.activityMs;
+DEFAULT_NATIVE_LEASE_SETTINGS.tileBudgetMb;
 DEFAULT_NATIVE_LEASE_SETTINGS.admission;
 //#endregion
 //#region src/custom-actions.ts

package/dist/export-alexa.addon.mjs

@@ -23857,13 +23857,24 @@ method(object({
 /** Playback-speed multiplier for the render (1 = realtime). */
 var ExportSpeedSchema = number().min(.25).max(32);
 /**
-* One dense interval, in SECONDS FROM THE EXPORT'S OWN `fromMs`.
+* One dense interval, in WALL-CLOCK SECONDS FROM THE EXPORT'S OWN `fromMs`.
 *
-* Relative and not absolute epoch on purpose: the renderer's frame-select
-* expression sees ffmpeg's `t`, which starts at 0 for the export's source
-* playlist. Handing it absolute epochs would make every call site responsible
-* for the same subtraction, and the one that forgot would emit a filter that
-* selects nothing — silently, as a uniform timelapse.
+* **Wall clock, not ffmpeg's `t`** — and the recorder translates. A caller
+* derives these bounds from things that happened at a TIME (a track's
+* `firstSeen`), while `t` runs over the source playlist: the concatenation of
+* every segment present for the range, with each recording GAP removed. The
+* two agree only on a window that recorded without one interruption, and only
+* the render side knows the segments, so the translation lives there
+* (`export-dense-map.ts`, addon-pipeline).
+*
+* It was not always so. These seconds were fed to `between(t,…)` verbatim, and
+* on a 10 h window holding 29,393 s of footage every range landed late by the
+* gap accumulated before it — up to 6,607 s, well past EOF. Nothing matched,
+* the video was a uniform timelapse, and the log line reported the five ranges
+* that had been ASKED for (2026-08-13, export `57d14363`, camera 615).
+*
+* Relative and not absolute epoch, because an absolute epoch would make every
+* call site responsible for the same subtraction.
 */
 var ExportDenseRangeSchema = object({
 	fromSec: number().nonnegative(),
@@ -30993,25 +31004,32 @@ object({
 var NativeLeaseAdmissionSchema = _enum(["all", "inferred"]);
 object({
 	/**
-	* How long a retained native frame is served before it counts as a miss.
+	* How many delivered frames the worker HOLDS at once, waiting for each one's
+	* detection result.
 	*
-	* Must cover the FULL late-crop horizon: detection inference + the
-	* cross-process inference-result hop to hub post-analysis + tracking + the
-	* tRPC crop round-trip back. Below ~500 ms the busiest cameras' subject crops
-	* outrun it and fall back to the ≤640 detection frame; above ~3 s the resident
-	* RAM per busy camera grows linearly with no measured hit-rate gain.
+	* This replaced a TTL on 2026-08-13, and the replacement is the whole point:
+	* a time window was never related to the event the pixels were waiting for.
+	* A held frame now lives from delivery until the runner has its `FrameResult`
+	* — at which moment the runner cuts the subject tiles it actually wanted and
+	* releases the frame. The bound exists only so a runner that stops answering
+	* cannot pin RAM: above it the OLDEST held frame is dropped and counted.
+	*
+	* Sizing: the steady state is `inferenceLatency × deliveredFps`, measured at
+	* 40-160 ms × ≤25 fps = 1-4 frames. The default leaves headroom for a hiccup
+	* without ever approaching the old resident set (43 frames × 24.9 MB at 4K).
+	* Raising it does not buy hit rate — it buys tolerance for a slow runner, and
+	* `holdOverflow` on the metrics line is what says you need it.
 	*/
-	ttlMs: number().int().min(250).max(1e4),
+	holdFrames: number().int().min(1).max(64),
 	/**
 	* Hard per-decode-worker RAM ceiling for retained native frames, in MB.
 	*
-	* Intended as a SAFETY ceiling with the TTL as the effective cap — but check
-	* which one is actually binding before reasoning from that. At the shipped
-	* 1024 MB and a 2 800 ms TTL, a 4K camera hits the CEILING first (~43 frames
-	* at ~24 MB each) and the TTL never gets to expire anything; `leaseMb` /
-	* `leaseFrames` on the metrics line say which. When the ceiling binds, a
-	* change that admits fewer frames buys retention WINDOW at constant RAM
-	* rather than giving RAM back — lower this knob if RAM is what you wanted.
+	* Since 2026-08-13 this is a SAFETY ceiling and nothing else: `holdFrames`
+	* is what decides how much is held, and the ceiling is the number above which
+	* something is wrong. Before that it was the effective cap — at 1024 MB with
+	* a 2 800 ms TTL a 4K camera sat pinned at `leaseMb:1020, leaseFrames:43`
+	* with the TTL expiring nothing, which is exactly the confusion the hold
+	* removes. `leaseMb` / `leaseFrames` still say what is resident.
 	* `0` DISABLES the lease entirely and falls the worker back to the tiny
 	* leak-prone GPU surface ring (~85% crop miss; that is what the lease exists
 	* to replace).
@@ -31037,22 +31055,45 @@ object({
 	* there is the signal that some caller names frames outside the inference set
 	* and that this must go back to `all`.
 	*/
-	admission: NativeLeaseAdmissionSchema
+	admission: NativeLeaseAdmissionSchema,
+	/**
+	* RAM ceiling per decode worker, in MB, for the SUBJECT TILES — the
+	* compressed native crops the worker cuts at the moment a frame's detection
+	* result arrives, and keeps long after the frame itself is freed.
+	*
+	* This is the knob that replaced the old retention window, and it buys about
+	* three orders of magnitude more of it: a tile is one subject at native
+	* resolution, JPEG-encoded (~60-120 KB on a 4K person), against ~24.9 MB for
+	* the frame it was cut from. A frame on which nothing was detected costs
+	* nothing at all, which is the real change — the old lease paid per FRAME and
+	* was interrogated per SUBJECT.
+	*
+	* `0` DISABLES tiles, leaving only the hold window and the ≤640 RAM
+	* fallback — i.e. the pre-2026-08-13 miss profile. Set it there only to
+	* reproduce that.
+	*/
+	tileBudgetMb: number().int().min(0).max(1024)
 });
 /**
-* The values in force when the operator has set nothing — byte-for-byte the
-* constants the decode worker shipped with as env-var defaults, so making these
-* settings changed no behaviour on the day it landed.
+* The values in force when the operator has set nothing.
+*
+* `budgetMb` stays at 1024 on the day the hold landed, deliberately: it stopped
+* being the retention window and became the OOM ceiling, and lowering a ceiling
+* in the same change that redefines it would make a regression and a retune
+* indistinguishable. Cut it once `tileHits` / `holdOverflow` have been read on
+* live traffic.
 */
 var DEFAULT_NATIVE_LEASE_SETTINGS = {
-	ttlMs: 1200,
+	holdFrames: 8,
 	budgetMb: 1024,
 	activityMs: 15e3,
+	tileBudgetMb: 64,
 	admission: "inferred"
 };
-DEFAULT_NATIVE_LEASE_SETTINGS.ttlMs;
+DEFAULT_NATIVE_LEASE_SETTINGS.holdFrames;
 DEFAULT_NATIVE_LEASE_SETTINGS.budgetMb;
 DEFAULT_NATIVE_LEASE_SETTINGS.activityMs;
+DEFAULT_NATIVE_LEASE_SETTINGS.tileBudgetMb;
 DEFAULT_NATIVE_LEASE_SETTINGS.admission;
 //#endregion
 //#region src/custom-actions.ts

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@camstack/addon-export-alexa",
-  "version": "1.2.18",
+  "version": "1.2.19",
   "description": "Alexa Smart Home Skill export — hub-side OAuth provider + directive handler. The companion AWS Lambda forwards Alexa directives to this addon's /addon/export-alexa/directive endpoint.",
   "keywords": [
     "camstack",