Skip to content

Phase 4: the card and the clip - #5

Draft
walkertbrown wants to merge 42 commits into
mainfrom
phase-4-card-and-clip
Draft

walkertbrown wants to merge 42 commits into
mainfrom
phase-4-card-and-clip

Conversation

@walkertbrown

Copy link
Copy Markdown
Owner

Phase 4 of approved/plan.md: the card and the clip.

Draft. Opened now so CI runs on every commit. Phase 4 isn't handed off yet.

On this branch but not approved yet (items 99, 100, 102, 103 wait on Walker's yes):

  • the referee bot: frames from referee_frames, three lamps, doubt with three whites, staging test answers
  • REFEREE_BOT stays off. With it off, only a staging test answer sends a close call to the bot.

Not built: item 98 (told not to), and item 101 (waits on LookieBot's look and Walker's yes).

🤖 Generated with Claude Code

https://claude.ai/code/session_01AJxnNf9Ug6uxYGyNkrRYsx

walkertbrown and others added 30 commits November 15, 2026 20:30
- An upload step for an attempt id that doesn't exist, isn't a uuid, or
  isn't hers answers 404 "no such attempt", the same either way. A bad id
  answered 500 before: it reached Postgres, which threw. This covers
  upload, uploaded and upload-failed. A DB test checks every step.
- samples.json's comment: the saved frame is used exactly (item 95), not
  "within one frame".
- The worker test's comment: a sample's frame route uses the points at
  exactly its saved frame; her own lifts use the nearest.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AJxnNf9Ug6uxYGyNkrRYsx
- web/lib/card: the card's words and numbers, the Story (1080 x 1920) and
  Post (1080 x 1350) drawn to the sample's .meetcard times 1080/390, the
  lit lamps as SVG with the sample's glows and bloom, and the heaviest bar
  with the app's plate split.
- GET /api/card/<token>.png?f=story|post, public by its token; 404 with no
  total or no card. The card token is made on first need.
- Static card fonts cut with fontTools (scripts/cut-card-fonts.py).
- 5.28: the total's Primary is "See my meet card" (bombed out keeps
  "See my lifts").
- Tests: the words, Home cards, both PNG sizes, and the footer staying
  inside the card for the longest plate lines.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AJxnNf9Ug6uxYGyNkrRYsx
…on /healthz

- attempts.clip_shared_at (item 99, waiting for Walker's yes): set by "Share" or
  "Copy my card link" on 5.30's Clip segment; the card link then shows the clip.
- reviews.by_bot and lamp_notes, and referee_jobs (3.4d, item 88).
- /healthz's R2 part HEADs every clip in samples.json, cached 10 minutes (item
  100, waiting for Walker's yes). Production's close-call clip was missing.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AJxnNf9Ug6uxYGyNkrRYsx
- /me/card (5.30): Story, Post, Clip and Print segments; Share (Web Share
  with the file, or a download), Save to my phone, Copy my card link; the
  honest line under the card. On Clip, Share or Copy puts this clip on her
  card link the first time (item 99, on the branch until approved).
- /me/card/print: the US Letter certificate (look 9), with the attempt sheet.
- /card/<token> (5.5): the story card, "Watch her lift" when the clip is on
  her card link, the body, the buttons, og:image and og:title.
- /api/card/<token>.mp4 streams the card clip with Range; ?download=1 saves
  either file. /api/me/card/share-clip reads and shares the clip.
- 5.4: the corner page shows her story card once she totals, June's too.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AJxnNf9Ug6uxYGyNkrRYsx
When her total lands (judge, bar tracker or referee lights), her card's lift
gets a judge_queue 'clip' row. The dispatcher starts a render_clip run (c-
worker ids, at most 2) only when no poses row waits. The worker draws look 9
with Pillow: her attempt (mirrored like the stage, except sample clips), the
lamps lighting at 330 ms with thocks, a 2 s hold with the marks, a 400 ms fade
to her story card held 3 s. It encodes a 1080x1920 H.264/AAC MP4 of 15 s or
less with the bundled ffmpeg and PUTs it to R2. A failure is tried once more,
then 'failed'.

Tests: worker/tests/test_clip.py (timing, thocks, layout, a real render),
tests/db/clip.test.ts (queue once, move on a referee change, poses first,
ready/requeue/failed, leases, the dispatcher), tests/unit/clip-window.test.ts.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AJxnNf9Ug6uxYGyNkrRYsx
…attempt

The card's line, its clip and the clip queue now share one pick, carried on the total.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AJxnNf9Ug6uxYGyNkrRYsx
After two failures her Clip segment shows "Your clip couldn't be made." with ghost "Try again"
in the clip's 9:16 frame on the stage (look 5.15, never red). Try again queues it once more
for one try; after that it's "We've told the team." and /ops ERRORS counts "Clips that
failed". FAULT=clip (staging) fails clips on purpose through the real path. Not approved
yet: waits for Walker's yes with items 96-103.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AJxnNf9Ug6uxYGyNkrRYsx
… yet)

- worker: referee_frames picks and stamps up to 60 frames (512 wide, time and
  last command burned in) and posts them back
- web: referee_jobs flow: enqueue, frames job, three lamps (left streamed
  first, head and right on its first event), second ask, doubt with three
  whites, lease and sweep, cost
- staging test answers fake only what Anthropic sends back
- REFEREE_BOT stays off; with it off only a staging test answer sends a job
- dispatcher starts referee_frames runs; instrumentation starts the bot loop
- migration 0009: frames tries, next try, ops_settings
- tests: DB flow for every test answer, lamps unit tests, worker frames

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AJxnNf9Ug6uxYGyNkrRYsx
/ops gets a REFEREE BOT section (on/off, key, waiting, median frames
ready to lights over the last hour, decided this hour, spend against
the monthly limit) and "Referee bot not running: <reason>" in red when
the bot is on with no key or doubted in the last 30 minutes. Staging
gets the "Referee bot test answers" choice for the next job (3.4d
step 7).

/healthz adds a `watch` (oldest queued pose age, the bot's not-running
reason and for how long), kept out of `ok`. The uptime check opens
"Lights are slow" (plan 2.4) and "Referee bot not running" issues past
10 minutes, and closes them when they clear.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AJxnNf9Ug6uxYGyNkrRYsx
Item 88: once REFEREE_BOT is on, close calls and benches say "the
referee bot" and "usually within a few minutes" (the stage, the bench,
the corner, the nine-attempt row, the lamps' label, the total's wait,
Try a lift's close call, the FAQ's privacy answer with Anthropic, and
the demo's link). While it's off (3.4d step 14) a person decides in
/referee, so the phase 3 words stay. A lift the bot couldn't look at
shows three whites with "The referee bot couldn't look at this one, so
it counts." The root layout reads the switch and hands it to the
browser's screens.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AJxnNf9Ug6uxYGyNkrRYsx
…ge the lights, the frames

Item 88 (plan 5.33, 5.36, look 5.14d), behind REFEREE_BOT:
- /referee is "Referee bot": "<n> deciding · <m> decided today", with Lights
  (the ones that stand) and Decided. Decided rows stay while their board isn't
  final, and today's stay.
- One close call shows the bot's lamp board above the toggles, which start at
  its lights. "Change the lights" is off until one differs; it's saved with the
  referee's id and stands. The page is read only once the board is final.
- The demo's bot page: the board lights left, head, right 330 ms apart once
  the lights are in, then the clip and 6 of the frames it read, with the judged
  frame ringed.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AJxnNf9Ug6uxYGyNkrRYsx
The sample route HEADs the sample before copying it. A missing one (or a
failed copy) answers 503, which the phone shows as the refused-upload
sheet, and the attempt and the demo count are left as they were.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AJxnNf9Ug6uxYGyNkrRYsx
The worker's referee-frames job now calls referee_frames.make_frames, which
picks, grabs, stamps and encodes an attempt's frames from a clip on disk
(running the bench's poses when asked). tests/referee-bot/recut.py will call
the same function, so the bot's test sees exactly the product's frames
(3.4d step 11). Behaviour is unchanged; two tests cover the shared path.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AJxnNf9Ug6uxYGyNkrRYsx
…ger clips

make_frames takes an optional window (ms of the clip) and treats that stretch
as the attempt, every time in and out counted from its start. The worker's
job never passes one; tests/referee-bot/recut.py will, to cut each by-eye
attempt from its source clip (3.4d step 11).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AJxnNf9Ug6uxYGyNkrRYsx
…ame code

tests/referee-bot/recut.py reads the set file and runs referee_frames.make_frames on each
attempt's stretch of its clip, so the test's frames are exactly the product's: picks, stamps,
size, quality and the bench's poses. frames.json records each frame's time and SHA-256 and the
set file's SHA-256; JPEGs are written only for clips the repo may hold.

Adds the Commons bench demonstration clip (CC BY 3.0, FitnessScape) to tests/clips with its
SOURCES.md row and by-eye calls: 3 bottoms, each "either".

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AJxnNf9Ug6uxYGyNkrRYsx
Plan 3.4d step 11. Each attempt in the set goes through bot.ts's
attemptWords and buildPrefix, lamps.ts's askLamps and makeClient, the
same code the worker runs. Every frame is read from recut.py's JPEGs and
must match frames.json's SHA-256 before it is sent, and frames.json must
have been cut from this exact set file. Held-back attempts (holdout.json)
are skipped unless --holdout, which runs only them. Each result records
the set file's and prompt.md's SHA-256, the model, the commit, the cost,
and for each bench the lines the bot was sent next to the eye's pause.
--test runs the whole path on the staging test answers, with no key.

The bot test scripts join the typecheck.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AJxnNf9Ug6uxYGyNkrRYsx
framesJob took the first command key starting with "press" as the bench's
Press!. A lost track (3.4b) writes `press_late_why: "lost"` next to
`press_late`, and that key starts with "press" too, so press_ms could be
the word "lost". It only worked because Postgres orders jsonb keys
shorter first. benchBlock now takes `press`, else `press_late`, and only
number values. Unit tests cover both key orders.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AJxnNf9Ug6uxYGyNkrRYsx
A phone that lost the bar has no path at Press!, so neither the pause nor
a sink after it can be known. measureBar measured across the gap when the
tracker found the bar again. On press_late_why "lost" it now measures
neither. The referee's lines and the bot read "Pause: not measured. The
phone lost the bar.", with no "Press! while sinking", even for measures
stored before this rule. Unit tests cover both.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AJxnNf9Ug6uxYGyNkrRYsx
build-set.ts makes every attempt through the product's own code:
takeBar's checkCommandTimes and measureBar for the bench, MeetLift and
judgeAttempt for the squat, deadlift and press sets. As ruled on Nov 15:
Press! 0.2 s before the eye's pause end (never before the stop) in cut 1,
a Rack! past the cut's end one server frame inside it, and the build
stops if any bench would go out as command_times.

score.py recounts every line of step 11 from the committed answers
alone, with its own tests in CI. run.ts gets --all for the final runs.
calls.json and holdout.json wait for Walker's yes.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AJxnNf9Ug6uxYGyNkrRYsx
…(3.4d step 1, item 107)

With no Press! the touch is the wrists' lowest point between Start! and
Rack! (or the end), and the lockout their top after it, with the poses
looked for all the way to Rack!. With no usable poses, the frames go
0.4 s either side of the saved bar path's lowest point (bar_low_ms,
read by framesJob from attempts.barKey); with no bar path either, from
Start! + 0.5 s for 1.5 s. Every time, plus from 0.6 s before Rack!.

Tests check where the frames fall, each fallback, the bar path reaching
the worker, and a real no-Press! rep of the bench demo getting 20 or
more. The set's 21 no-Press! reds go from 5-9 frames to 15-30.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AJxnNf9Ug6uxYGyNkrRYsx
…w (3.4d step 1)

The frames job sends the saved bar path on every bench. worker/bench_path.py
finds the touch (the path's lowest point to Press! or Rack!) and the lockout
(its top after, to Press! + 3 s or Rack!), using the path when it isn't lost
before Press! and has a point at confidence >= 0.5 on 80% of the frames to
the touch. A path lost partway joins the wrists after the loss; else the
wrists alone when seen on half the frames; else fixed windows. Poses run
only when the path isn't enough.

The lost-track cut sends the whole saved path (the product saves the
server's own track); measureBar and the lines still stop at the loss.

Tests: each way in, with and without Press!; a real trace thinned to 90%
and 70%; points under 0.5 confidence counted as missing. test_touch.py
checks every path touch is within 0.2 s of the reference (skips until
calls.json is committed). README: per-bench way and distance table.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AJxnNf9Ug6uxYGyNkrRYsx
…ce named (3.4d step 1)

test_touch.py now checks that the frames sent hold every frame within 0.2 s of the reference's
lowest point, on every bench with a reference (The Playner's way 1, Nov 16). It passes on every
bench, and fails on 15 benches with the frames cut before the bar-path rule.

The README names IA0P rep 2 (both cuts) as more than 0.4 s off: the phone's plate trace froze for
1.5 s and rose 0.4-0.5 s late. It says why 6K8v rep 3 takes the wrists (the committed trace ends
early), that the plan's readings are now the plan's own, and that the set's paths are the phone's
while a meet measures the server's re-track.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AJxnNf9Ug6uxYGyNkrRYsx
…the new stuck line (3.4b, 5.33)

- stuckMs 800, counted as frames x frame time; IA0P's server reps 1 and 3 pinned, 11 frames not caught.
- A still stretch that starts after Press! (or the lowest seen point) and stays within the back-up band is a
  lockout hold, not a stick (hIuj's lost cut).
- The stuck line: "The phone lost the bar around the pause, so its Press! may have come early. If she pressed
  on or after the call, don't give a red for the pause."
- The set: IA0P rep 3 "Press! inside a stuck stretch" (white on the pause), held back with IA0P.
- The server's tracks (tests/bench-commands/server.py, data/server/), the stuck count on /ops.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AJxnNf9Ug6uxYGyNkrRYsx
- score.ts plays every rep of the gym sets with ALL_REPS=1. Reps 2 and on
  take their moments from data/calls-run1.json with the 10% rule and a
  stand-in hold: Start! on the stillest stretch before the descent, Rack! on
  the highest stretch after the back-up, Press! as on rep 1.
- bar-caller.ts takes an optional frozenPressMs (off in the product's
  rules). When x and y stayed within 0.002 through Press!'s 0.5 s, Press!
  waits until the bar has been still that long. Frame by frame, with no
  look-ahead.
- results/commands-all-reps.json holds every line and row with the rule
  off and at 1.0 s. The tests pin IA0P rep 3 as the only early Press! with
  it off, and the three Press! lines passing at 1.0 s.
- results/commands.json only gains each row's rep number.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AJxnNf9Ug6uxYGyNkrRYsx
…time

- measureBar's median and the worker's are over 100 ms either side, plus 5 ms
  slack (smoothMs), not 3 points, so the live tick's repeats of one camera
  frame don't undo it (item 110, The Playner, Nov 20, 8:47 AM).
- The 0.99 stuck rule's window is full when its oldest point is within the
  path's median gap of its start, and its time is 0.4 s before Start! and
  1.5 s after it (freezeBeforeMs, freezeAfterMs; item 111, 9:02 AM).
- Item 109 in the product's rules: frozenPressMs 1000.
- The commands test runs every rep at the grid, the 16.7 ms tick and each
  frame fed 3 times: lines 1-4 on every rep; no command in the clip moves by
  more than two of the clip's own frames (holds recorded, not counted); and
  on each rep's whole phone path, with item 108's guard, no real pause of
  0.5 s or more reads "not measured" because the live stuck rule marked it
  lost (NerdCider, 9:01 AM). That check fails at 0.4 s (IA0P reps 1, 2).
- results/commands.json and commands-all-reps.json regenerated; the all-reps
  file is labelled as the stand-in (two reps of one clip).
- The referee-frames trace test: the touch on the 100 ms median is at the
  middle of the flat pause.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AJxnNf9Ug6uxYGyNkrRYsx
…t 15-60 fps

One bad camera frame smoothed away at each rate, fed once or 3 times; the
stuck rule fires 0.4 s before Start! and 1.5 s after it, within one frame,
with even and uneven frame times (items 110, 111).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AJxnNf9Ug6uxYGyNkrRYsx
The Playner's 9:11 AM sweep (item 111). freezeNearTap (unset in
shared/rules.json, so the product is unchanged) counts a stuck stretch
before Start! only while the track is within D of the tap, measured as
leftRack is. sweep-near-tap.ts runs today's rule, D = 5, 6 and 8%, and
the rule off before Start!, on every rep at the three cadences, with
the scores as recorded and lifted to 0.995, plus m-k8 with the tap at
9.0 s. score.ts gains lift, tapAt and only options and a
stuckBeforeStart field; the results files gain only that field.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AJxnNf9Ug6uxYGyNkrRYsx
The Playner's ruling (Nov 20, 9:23 AM): freezeNearTapPct 8 in shared/rules.json,
measured as leftRack is; after Start! it stays 1.5 s anywhere.

Tests:
- bar-timing: a still bar 7.5% from the tap goes stuck before Start!, 8.5% never
  does, up and to the side.
- bench-commands: the 13 racked clips' real holds with every score lifted to
  0.995 are never lost before Start! at the grid, the tick or fed 3 times, but
  ak3m3 at the tick (accepted); m-k8's racked bar never gets Start!, and with
  her tap at 9.0 s her hold does, as recorded and lifted, at all three.
- The cadence line's one-run list is now 8qzW's Rack! only: VM_y reps 2 and 4
  get Start! at the live cadences.

Results regenerated: only VM_y rep 1 (voided at 7.93 s, not by the stuck rule)
and the stand-in's m-k8-r3 (Start! again) move.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AJxnNf9Ug6uxYGyNkrRYsx
walkertbrown and others added 12 commits November 20, 2026 09:30
Item 111 (plan 2.4, 3.4b): under the bench counts, every lifter with two or
more lost_bar_void rows in one meet, over the same filmings as "No Start!,
filmed again" (the demo and test meets included, never a sample or a Home
press set), with her count.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AJxnNf9Ug6uxYGyNkrRYsx
…its it

- The caller names each void's cause: no_start_5s, stuck or low_score
  (3.4b, "The void has three causes"). low_score has its own key,
  lowScoreBeforeMs: 1000 (the Playner, 9:39 AM), so the stuck window
  never moves it.
- The phone posts the void as soon as it's voided, with { why }; the
  server stores it (anything else is the old no_start) and answers n,
  its count since the attempt's last upload_started. After "Rack!"'s
  900 ms the phone uses the server's count, or its own if none came.
- The sheet follows this void's cause, first and second in a row, with
  the plan's words; only the stuck second sheet goes back to "Find your
  spot". On no_start_5s nothing is said after "Rack!".
- The sheet's text scrolls above the pinned button on short screens, with
  the 1px rule line only while text runs under it (5.14c).
- /ops: "No Start! twice or more in a meet" shows each lifter's causes.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AJxnNf9Ug6uxYGyNkrRYsx
… counts it

- MeetLift passes the caller's `held` on its Press! call, and the stage
  writes `command_times.press_held: true` with it (3.4b, item 109).
- The server takes press_held only as `true`, and only with a Press!
  key; anything else is a command_times close call, like a broken time.
- /ops "Bench, phone and referees": "Press! held for a frozen track",
  over the same bench filmings as the tracker's other rates, with no
  10% line (2.4).
- Unit tests: a frozen pause gets Press! 0.5 s after a still one that
  wobbles, marked held and never late; off, nothing is held.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AJxnNf9Ug6uxYGyNkrRYsx
…scape screen

At 300 px tall and under, the landscape panel scrolled and took the button
with it, and the text box never overflowed, so its rule line never showed
(LookieBot's early read, Nov 20, 9:46 AM). On the void sheet the panel is
now held to the screen: the title and Body scroll above the button, which
stays 16 px from the bottom. At 568 x 280, the button sits at 208-264 and
the text scrolls 22 px; 320 px and up are unchanged. A class, not :has(),
for older phones.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AJxnNf9Ug6uxYGyNkrRYsx
…e's path

The set's benches are now measured on the phone's own path, as takeBar
does (item 110): the trace in the cut, item 108's guard on it, then
measureBar. The frames still find the touch on the server's guarded
track, so test_touch.py's numbers are unchanged.

pauses.ts measures every bench's Pause line at 30, 20 and 15 fps cameras
(frames dropped from the trace), each on the grid, at the live tick and
fed 3 times, unplanted and with one bad frame planted in each real pause,
0.05 lower and higher: 27 runs. A changed trace never moves a cut.

test_pause.py (3.4d step 11) fails if a pause of 0.5 s or more by eye is
sent under 0.3 s, or one under 0.3 s is sent 0.5 s or more; "not
measured" never fails. Every unplanted run passes, and the 14 benches
with a Press! on a pause of 0.5 s or more are each sent a number. A
planted line must stay within one camera frame of its unplanted run:
19 planted runs miss today (ak3m3 at 15 fps, z838 rep 1, IA0P reps 2 and
3, uGCh at the tick). They are listed in KNOWN and printed every run, for
The Playner to rule on; a new miss fails, and so does a known one that
stops missing. 8qzW rep 1 is recorded, not checked, its number printed.

Also: score.ts imported CommandTimes from bar-measure, which doesn't
export it; the referee-bot typecheck now covers it through pauses.ts.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AJxnNf9Ug6uxYGyNkrRYsx
…go to the bot as recorded rows

- stuckStretches (item 108's guard): one camera frame outside the 0.002 box, with the next
  back inside the same box, is part of the stretch and marked with it, and never widens the
  box. One camera frame is a run of identical points lasting no more than a camera frame,
  so the live tick's repeats count once (NerdCider, Nov 20, 10:00 AM; 3.4b). Two off
  frames end it.
- IA0P rep 3's server stretch runs to 17.67 s (was 17.53), through one off frame.
- test_pause.py: the eye check is hard on planted runs; moved lines are printed; KNOWN holds
  only ak3m3 at 15 fps, 0.05 lower; IA0P rep 2 and IA0P-r3-stuck read "not measured" on
  every camera, tick and plant.
- build-set.ts: ak3m3-planted and 8qzW-r1-planted, cut 1 with the planted 15 fps path, as
  recorded rows. score.py counts them in no line and prints their lamps beside AMih, ak3m3
  and 83WA-r2's; a red for the pause on any goes to The Playner.
- commands-all-reps.json: xXFj rep 3's held still stretch now reads stuck (pause 0.86 to
  not measured).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AJxnNf9Ug6uxYGyNkrRYsx
…third check on item 113)

A real pause of 0.5 s or more by eye, measured on the recorded trace, must be measured on
every camera, tick and plant. KrDo (eye 1.07 s) at 15 fps on the grid and the tick isn't
since the guard carries one off frame: her still bar wobbles 0.0021 x 0.0026 and each
single frame off the 0.002 box is carried, so 3.97-4.70 s reads stuck. Listed in LOST_REAL
for The Playner to rule on; printed on every run, and the test fails if it changes.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AJxnNf9Ug6uxYGyNkrRYsx
…ckMs

3.4b, "Only on a frozen track's motion" (The Playner, Nov 20, 10:13 and 10:16 AM;
NerdCider's two holes, 10:15 AM):
- stuckStretches carries an off frame only when the median step from the
  stretch's start up to the off frame, between points a camera frame apart in
  time, is under stuckCarryStep (0.0002 of the frame, shared/rules.json).
  Otherwise the off frame ends the stretch, as before.
- A stretch is stuck when frames x frame time reaches stuckMs less 0.5 ms.
- Unit tests: a frozen track carries, a real still bar (0.0007 a frame) and
  0.003 flips don't, fed 2-4 times, nudged repeats, frozen-then-moving, and a
  stretch at exactly 800 ms at 30, 20 and 15 fps on every tick.
- test_pause.py: LOST_REAL is empty; measured real pauses (KrDo's among them)
  are pinned measured on all 27 runs.
- results/commands.json and commands-all-reps.json regenerated: xXFj rep 3 is
  measured again (0.86 s).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AJxnNf9Ug6uxYGyNkrRYsx
medStep answered 0 when there was no step, so an off frame that came as a
stretch's second camera frame was carried whatever the bar did (VibeSearch,
Nov 20, 10:25 AM). It now answers Infinity: no step, no carry. Every Pause
line on all 27 runs is the same as at ce873ab.

Also the unit test The Playner asked for on item 114's fallback: a frozen
track exactly identical for 1 s is stuck at 30, 20 and 15 fps, one point per
camera frame and on the refresh tick.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AJxnNf9Ug6uxYGyNkrRYsx
- The phone's stage stamps each tap with its time in the recording: 0 for
  the first tap (just before "Bar is loaded." starts it), the re-tap's own
  time after. A second tap is the last tap's point and time.
- The server reads t when present (0 or more), checks it sits inside the
  clip ('outside_clip' otherwise), and a broken t is a command_times close
  call. An old phone's tap has none.
- The worker's tracker starts at the first 15 fps frame at or after t.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AJxnNf9Ug6uxYGyNkrRYsx
One frame, the off frame, then 0.73 s still at 30 fps (the 15 fps case was
already in): not stuck, on the grid and fed 2-4 times.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AJxnNf9Ug6uxYGyNkrRYsx
…rame

The Playner, Nov 20, 10:28 AM (NerdCider, 10:27 AM): one step is nearly as
loose as none, so with fewer than 2 steps nothing is carried. Steps are taken
at most one per camera frame in time (a step's point at least 0.9 of a camera
frame after the last step's), so a frame fed 2-4 times is still one step:
otherwise one real step fed twice counted as 2.

NerdCider's test: a real bar whose off frame is its third camera frame, with
the one step before it under 0.0002, isn't carried, at 15 and 30 fps on the
grid and fed 2-4 times. Every Pause line on all 27 runs is the same as at
ce873ab.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AJxnNf9Ug6uxYGyNkrRYsx

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant