SPC: the sampled chart drawn, and a click opens the five bottles behind the dot

[kalwei]

What

Quality → SPC draws a sampled characteristic, and a click on a dot opens the
readings behind it.

Until now the screen drew one kind of control chart. A characteristic with a
sampling plan (#145) had its arithmetic, its samples and its verdict, and a
sentence saying the screen could not draw it. It draws it now:

• X-bar and R. The mean of each sample above with X̿ and X̿ ± A2·R̄,
  the spread inside each sample below with R̄, D4·R̄ and D3·R̄, the
  specification limits where there are any, one dot per sample on each half.
  It is the IMR layout from SPC: the moving-range half of the chart, drawn — an IMR chart #144, not a third one: the same two halves in one
  <svg>, so one export takes both; the same selected-dot ring; the same
  promise that the chart does not move when the panel opens.
• A click opens the n readings. On an individuals chart a point is a
  reading and the panel opens that reading. On a sampled chart a point is an
  average, so the panel opens the five bottles it is the average of — each
  one's value, gauge, who took it, how far it sat from the sample's own mean,
  and which of them is furthest out. That is the question an X-bar chart
  cannot answer on its own: five bottles high is a filler setting, one bottle
  high is a nozzle.
• The window is the stretch those readings span, plus the minutes before,
  and every one of their stamps is marked on every trend — so which bottle was
  measured while the nozzle pressure was falling is visible rather than
  inferred. Everything else is Click a point on the SPC chart and see why it is there — no model, just the records #143's blocks asked of a sample: where it sat on
  both halves and what fired, the gauge and its calibration, the machine's
  state and what it had just come out of, the stops, maintenance, findings, and
  each block's own coverage.
• GET /quality/spc/{material}/{characteristic}/sample/{sample_id} is the same
  read as an API, with the same before_minutes / after_minutes /
  neighbour_hours as the point read. Nothing in it is computed for the panel.

One envelope key was added, as the brief allows: a top-level range_chart
block on the chart read, beside moving_range — centre, upper, lower,
n, stable, verdict, flagged — so the lower half's words come from the
same place whichever kind of chart it is, with or without limits. It carries no
signals of its own: on a sampled chart rule 5 judges a sample and already
arrives in signals indexed by the same samples as rules 1–4, and a second
copy would have a reader counting every range firing twice. control.range_chart
(the #145 contract) is unchanged, key for key.

And a caption bug the lab found. On 2026-10-06 the lab's bottling plant,
with no samples yet, drew fill_height under the heading INDIVIDUALS AND
MOVING RANGE — 0 SAMPLES with the individuals legend and the individuals
explainer beneath it. The heading, the legend and both sentences now follow
kind whether or not there is anything to draw. Decision 0040 made the chart
type a property of the plan; a caption that does not follow it undoes that.

Why

Scott, 2026-10-06: "test if this type of graphing and functionality can apply
to a new type of SPC." It applies — the layout, the dot behaviour and the
panel carried over without a second design, and the one thing that genuinely
differs (a point is an average, so there are n readings and a window rather
than an instant) is two blocks and no new page.

How verified

• pytest: full suite green. 33 new unit tests for the sample dossier
  (test_a_sample_on_the_control_chart_carries_the_five_bottles_behind_it.py),
  the chart half of the sampled-screen tests rewritten, and 25 Playwright
  tests in Click a point on the SPC chart and see why it is there — no model, just the records #143's style
  (test_clicking_a_sample_on_the_spc_chart_opens_the_five_bottles.py): the
  dossier fetch held back 1.2 s through a route intercept, every assertion
  waiting on the panel's own data-sample — which the page writes last —
  rather than on the box that will hold it; four themes with the palette
  probe; an operator and a supervisor both open it. They key on identity: one
  dot is flagged and it is the sample after the changeover, the bottle that
  pulled its average is the 147.0 mm one.
• A scratch bottling plant from the lab pack at 60×, on 127.0.0.1:8765 /
  OPC 4871, stopped afterwards through fsmes plant … stop; the lab fleet's
  8090/8110/8120/4941/4942 and the demo's 9030/5433 were listening before and
  after. Sixteen samples of five bottles, limits appearing on their own at
  spc_min_points, X̿ 141.71, UCL 143.46, R̄ 3.04, D4·R̄ 6.42, verdict
  in control but not capable (Cpk 0.69). Clicked; the panel opened on the
  five readings with the furthest marked; added a four-at-a-time characteristic
  with no readings at all and the heading, legend and both sentences were the
  sampled ones.

(The image is in the diff at docs/plant/images/spc-xbar-and-r.png; the figures
above are the same run.)

The acceptance table from #143, extended

The bottling pack plants one fill-height story: the nozzle the changeover left
behind — for twenty-five line minutes after each changeover the sample means
sit about half a millimetre high while the ranges stay where they were.

Cause	Where it lands	Clicks
the nozzle the changeover left behind — is the sample findable?	No rule fires on it. Every post-changeover sample in the run sat at 141.90–142.22 and every other sample at 141.26–141.80 — six and nine, not one overlapping — but the control limits are X̿ ± 1.75 mm, because R̄ is 3.04 mm. A planted 0.5 mm shift against a 3 mm within-sample spread is a sixth of the limit width. A reader sees the saw-tooth by eye; nothing on the screen points at it	by eye, then 1
…and once the sample is open, is the changeover named?	No. The sample is taken at QI01, the inspection bench; the changeover is at FILL01. The machine block is the sample's own station's, so it shows the bench going idle between inspections, and the three trends are CameraLux, InspectMs, RejectPct. The filler is one machine upstream	never, today
the five readings, and which one pulled the average	first block, every time: value, distance from the sample's own mean, the furthest marked, with the sentence that one of them always is	1
whether it was the spread rather than the process	the lower half's own limit and verdict are in the rules block beside the upper half's	1
the gauge	HEIGHT-FILL-01, calibrated 2026-09-24, due 2027-03-23, resolution 0.1 mm against a 6 mm tolerance	1

So: not in one click, and the reason is the same one #143 found with the
washer. The panel answers what happened at this station, and the planted
fill-height cause is a station upstream — the link to the line's other stations
over the same window is on the board as its own item, and this PR does not
build it. The second finding is about the pack rather than the product: a
simulated bottle-to-bottle spread six times the shift it plants makes the
story invisible to any control chart, including a correct one.

Honest, about the screen

The panel is 4,114 px tall on a sample. That is not a regression — the same
plant's point panel is 4,140 px — but on a sample the first screenful is
now the readings table, so what the rules said starts below the fold. The
block that should be collapsed first is this station's process values
(1,122 px). Collapsing the lower blocks is the separate waiting item and is
deliberately not in here; it is needed.

Not in here

No model. No new tab, no new table, no migration. The upstream-station link and
the collapsing panel are both on the board. Count charts and quality-team limit
editing are out of scope. control.range_chart, the individuals chart and the
point dossier are unchanged.

The red cell on this PR, and the seventh commit that fixes it

Everything passed after the rebase except test (windows-latest, 3.12), and
the two attempts failed on two different tests, neither in a file this
branch touches. Both are now fixed here, because both block a strict main and
a third re-run would only have found a third one.

tests/test_the_simulated_floor_plants_causes_an_engineer_can_find.py::test_the_heights_after_a_changeover_sit_half_a_millimetre_high
asked one sample of five either side of a changeover to see the pack's 0.5 mm
offset: the planted lift measured 0.06 mm against a tolerance of 0.35 mm.
Five bottles held at one weight still stand about 0.36 mm apart, so the
standard error of that difference is about 0.23 mm — the assertion was roughly
one standard error wide. It is worse at night, where the pack widens the
bench's spread from 0.10 mm to 0.25 mm, which is why it failed repeatably:
checked out and run here it failed 12 times out of 12 on main at 3bdc0886
and on f86ced15 before it, because the clock made it a night shift. It now
tests what the pack actually promises, in the pack's own words — "a shift a
reader finds by splitting the means on the changeover, not one that trips a
control rule" — over the run: twelve samples a side, and the tolerance is
four standard errors of the difference worked out from the pack's own spreads
at the wider of the two shifts rather than a number chosen by hand. The shift
it is looking for is six standard errors out, so an offset that went missing
is still caught. Over ten seeds the answer landed 0.39–0.62 mm, at most 1.5
standard errors from 0.5.

tests/test_read_never_waits.py::test_a_login_answers_while_the_floor_screen_is_computing
asserted a one-second wall clock (1.06 s on main's own ubuntu-latest, 3.13
in run 37497028434; 1.41 s here on windows-latest, 3.12). I took the first
of the two options: the sign-in is now measured against the screen's own
measured time, not the clock — it has to answer before the screen is half
way through. That comparison only became true once the screen was genuinely
slow, and making it genuinely slow turned up the real bug in the test: the
refresh it was supposedly racing was never signed in, so it was a 401 that
opened no transaction and held nothing. The test has been measuring a sign-in
with nothing behind it since #80. Both clients now sign in, and the screen is
held open inside its own read transaction by a time.sleep, which hands the
GIL back — the other reason the old number drifted, since TestClient runs
the app in this process and a refresh that computes in Python starves the
sign-in thread regardless of any lock. The test asserts separately that the
screen really was held open, so the comparison has teeth.

Tests only in that commit; no behaviour changes, so no CHANGELOG entry.

[github-actions]

Thanks for the first pull request. Checks: every commit signed off (git commit -s), ruff and pytest green, and the house-rules checklist in the template. The maintainer reviews within a few days.