9.5: runtime tests for tap.python, in Max

CLAUDE.md

@@ -72,7 +72,9 @@ in the same binary by 9.3; its documentation (9.4) and runtime tests (9.5) are s
 - **`runtime-tests/`** — tests that run inside a real Max (plan 6.1): Cycling '74's max-test harness
   (submodule), patchers generated by `make_patchers.py` (edit the generator, never the patchers),
   fixtures (`python/maxtest_*.py`, copied into `python/` for a run) and `run.py`, which installs the
-  harness into Max's Packages, launches Max, drives it over OSC and reports. macOS, local only —
+  harness into Max's Packages, launches Max, drives it over OSC and reports — `tap.python~`'s tests
+  and `tap.python`'s (`tap.python.*`, plan 9.5; `tap.python.load` opens first, in a Max with Restore
+  Windows on Launch turned off for the run, so that `init/` must find the object). macOS, local only —
   `runtime-tests/README.md` has what the first runs taught about writing one. Two sessions run only
   when asked for: `soak` (plan 6.2, an hour) and `perf` (plan 6.3). `runtime-tests/spike/` is the
   record of the 9.0 spike for `tap.python` — its patchers, runners and Max's logs; the spike's class

docs/PRODUCTION-PLAN.md

@@ -828,7 +828,7 @@ row shipped what only a host platform could show.
   generator); a vignette and three tutorials with patchers for the Documentation window; an
   Extras overview named as the landing patcher; help patchers with tabs for both objects; CI
   checks for the hand-made XML and JSON.
-- [ ] **9.5 Runtime tests:** the `tap.python.*` patchers in `make_patchers.py`; `run.py` aware of
+- [x] **9.5 Runtime tests** (#50)**:** the `tap.python.*` patchers in `make_patchers.py`; `run.py` aware of
   the second object; the soak session gains a `metro`-driven `tap.python` under saves.
 - [ ] **9.6 The release session, Mac and Windows:** the whole suite on the Mac, the help patcher
   checked and re-saved, Max's page for `tap.python~` committed, the hand checks; the package by

docs/TAP-PYTHON-PLAN.md

@@ -8,7 +8,7 @@ change — and the plan to build it (Phase 9 of the production plan, which point
 Drafted 2026-10-02 against 1.0.0 and audited the same day (`AUDIT-TAP-PYTHON-PLAN.md`: two
 blockers, six major findings, nine minor; its addendum of 2026-10-09 records what 1.0.1 and 1.0.2
 changed). **Revised 2026-10-09 against 1.0.2** for every finding; the [revision record](#revision-record)
-at the end says what each changed. 9.0 to 9.4 are built — the spike ran in Max on a Mac
+at the end says what each changed. 9.0 to 9.5 are built — the spike ran in Max on a Mac
 and on Windows: its answers are under [9.0](#what-90-found), and the design below follows them. Tick the
 items below (with the PR) as they land, and keep the design current where a PR decides differently.
 
@@ -450,15 +450,30 @@ third was Windows-only.
   9.5's), which also shows that the wrapper passes a class's attributes and messages on only through
   `setvalue` and `applyvalues`. The ReadMe's message sentence is the perf script's, written when
   the tables are next measured (9.6). Checked in Max 9.1.5 on the Mac; Windows in 9.6.)*
-- [ ] **9.5 Runtime tests.** The patchers above in `make_patchers.py` (9.4 added `tap.python~.mc`,
+- [x] **9.5 Runtime tests** (#50). The patchers above in `make_patchers.py` (9.4 added `tap.python~.mc`,
   for the `init/` mapping of `mc.tap.python~`); `run.py` aware of the
   second object (the `init/` mapping in the installed package, the `--package` mode, the
   reference-page rule for `tap.python~`'s page only). What 9.0's runner learned: a fresh Max
   needs Restore Windows on Launch off (a restored window loads the binary first); patchers are
   opened with an "open document" event, never `max openfile` through the harness's OSC (it crashes
   Max); and Max's log drops a `[print]`'s name. The soak session (6.2) gains a
   `tap.python` driven by a `metro` while its file is saved every second: every output checked,
-  memory flat, the console counts exact.
+  memory flat, the console counts exact. *(As built: seven `tap.python.*` patchers — load (opened
+  first, in a Max with Restore Windows on Launch off for the run, so `init/` must find the object),
+  outputs (every row of the table), messages (attributes, `get<attr>` from the dumpout, the
+  forwarder), reload (an outlet added before the dumpout takes a scripted cord and carries; one
+  removed takes its cord; a broken save reported once), shared (one file for both kinds: both follow
+  a save, each kind's reserved name said once — `mode` by `tap.python~`, `dumpout` by `tap.python`,
+  as Max answers `dumpout` on neither but the class that registers it), threads (a `metro` every 5
+  ms, every output against its input, on the main, scheduler and audio threads) and faults. Outputs
+  are checked as `[tosymbol]` writes them — floats with four decimals, `symbol` kept before a
+  string — which a probe in Max found; the README records it. The `[print]`-free checks needed no
+  `max openfile`: run.py opens patchers through the harness's `test.master run`, which 6.1 already
+  used. `run.py` requires the package's `init/` lines, warns if the hand-written page changes, and
+  puts Max's preferences back. All 21 runtime tests pass in Max 9.1.5 on macOS; the hour-long soak
+  passed all 18 of its checks, the `tap.python` answering every message of a 10 ms metro correctly
+  for 60 minutes of saves: Max's memory 658 MB a minute in, 663 MB at the end (peak 675), 3601
+  executions of the file, Python's object count between 18,700 and 22,000 throughout.)*
 - [ ] **9.6 The release session, Mac and Windows.** Build, run the whole runtime suite on the Mac,
   check the help patcher and re-save it, commit the page Max writes for `tap.python~`, the hand
   checks the tests cannot make (`get<attr>` through the dumpout in a patcher, attrui on a

runtime-tests/README.md

@@ -1,7 +1,8 @@
 # Runtime tests in Max
 
 The unit tests (`core/tests/`, and the external's `_test.cpp` against min's mock kernel) never load
-the object into Max. These do: each is a patcher that Max runs, driven by Cycling '74's
+the objects into Max. These do — `tap.python~` and, since plan 9.5, `tap.python` (the
+`tap.python.*` patchers): each is a patcher that Max runs, driven by Cycling '74's
 [max-test](https://github.com/Cycling74/max-test) harness (MIT, the `max-test` submodule), checking
 what only a real Max shows — the file watcher, `getattr` on the Python-defined attributes, audio
 through the real signal chain, `poly~`, the console, and loading without a runtime. (Plan 6.1 and
@@ -41,26 +42,32 @@ runtime-tests/
 
 ## What the runner does
 
-1. Builds the harness's `oscar` extension if needed, and installs the harness into `Packages` as
+1. Checks the package: the external built from the current sources, a runtime, and the lines of
+   `init/tap.python.txt` without which Max finds neither `tap.python` nor `mc.tap.python~`. Then
+   backs up Max's preference files, turns off **Restore Windows on Launch** and moves aside the
+   workspaces Max would reopen (into `logs/`) — a window reopened at launch would load the binary
+   before the first test — and puts the preferences back when the run ends.
+2. Builds the harness's `oscar` extension if needed, and installs the harness into `Packages` as
    `max-test` — a **copy**, not a link, because Max loads a package's extensions only from a real
    folder. The copy is marked `INSTALLED-BY-PYTHONTAP.txt`, replaced on every run, and left
    installed afterwards (Max then loads `oscar` at every launch; delete the folder to uninstall).
    Another `max-test` already in `Packages` is refused, not replaced.
-2. Links `runtime-tests/` into `Packages` as `PythonTap-runtime-tests` (so Max finds the patchers
+3. Links `runtime-tests/` into `Packages` as `PythonTap-runtime-tests` (so Max finds the patchers
    by name), and copies `python/maxtest_*.py` into the package's `python/` folder (the object loads
    classes from there only). Both are removed afterwards.
-3. **Session `without-runtime`**: moves `support/` aside and starts Max. The object must load and
+4. **Session `without-runtime`**: moves `support/` aside and starts Max. The object must load and
    say the runtime is missing; then, with `support/` put back while Max runs, a new object must
    say to restart Max (on macOS the weak binding to libpython is fixed when the external loads).
    `support/` is always put back — and a run killed midway leaves it as `support.maxtest-aside/`,
    which the next run restores.
-4. **Session `main`**: starts Max again and runs every `*.maxtest.maxpat`.
-5. Only when asked for: **session `soak`** (plan 6.2) runs `soak/tap.python~.soak.maxpat` for an
+5. **Session `main`**: starts Max again and runs every `*.maxtest.maxpat` — `tap.python.load` first,
+   so that the fresh Max must find `tap.python` through the package's `init/` mapping (plan 9.0).
+6. Only when asked for: **session `soak`** (plan 6.2) runs `soak/tap.python~.soak.maxpat` for an
    hour, sampling Max's memory every minute, and writes a summary — memory, reloads, and what the
    patcher logged each minute — to `logs/soak.summary`; **session `perf`** (plan 6.3) runs
    `perf/tap.python~.perf.maxpat` and averages Max's CPU-meter readings into `logs/perf.json`,
    which `scripts/update-perf-docs.py --max` turns into the ReadMe's table.
-6. Reads the results from the harness's SQLite database (in the installed `max-test`) and prints
+7. Reads the results from the harness's SQLite database (in the installed `max-test`) and prints
    them. Max's standard output goes to `runtime-tests/logs/` — only min's own lines appear there;
    the console errors seen during a failing test are printed from its log instead.
 
@@ -97,13 +104,27 @@ first runs taught, built into the helpers:
   rate is not like for like, so the perf patcher measures at the device's own rate.
 - **A check that audio did not change passes vacuously if no audio ran**: pair each with a sample
   check (the soak's `phase-N-…-output`).
+- **`tap.python`'s output is checked as `[tosymbol]` writes it** (`Test.output_is()`): one symbol for
+  the whole message — `symbol C` (the word `symbol` kept), `1 0 1`, `note 60 100`, and floats with
+  four decimals, `0.5000` — held in a `[zl.reg]` until the check compares it with a quoted
+  `[sel "…"]`. A `bang` makes `[tosymbol]` repeat its last symbol, so a bang output is not checked
+  this way. An output that never came fails the check: the harness records an assertion that never
+  answers as a failure ("never returned results"). `Test.output_count()` and `Test.count_is()` check
+  how many messages an outlet sent — nothing, for `None`.
+- **Text logged with spaces breaks the harness's SQL** (its log line goes between double quotes, and
+  Max quotes a symbol with a space): log codes, not text — `[atoi]`, as the probe that found the
+  formats above did.
+- **Overdrive is `; max preempt 1`, Scheduler in Audio Interrupt `; dsp takeover 1`**; the threads
+  test reads both through `[adstatus]` first and puts them back at the end.
 
 The fixtures' docstrings say what each is for. Open a patcher in Max to watch a test run; its
 `test.*` objects only record anything when the harness opened it.
 
 ## Not covered here
 
-Some of the runbook's checks still need a person: that a `bool` attribute shows as a toggle in the
+`tap.python`'s once-per-session notice that Scheduler in Audio Interrupt runs Python on the audio
+thread is a warning, which `[error]` does not see: the glue test checks it, and the threads test only
+that the messages run there correctly. Some of the runbook's checks still need a person: that a `bool` attribute shows as a toggle in the
 inspector, that an `attrui` displaying a removed attribute looks right (the tests check only that
 nothing breaks), and whether the audio *device* drops out while a reload holds the GIL (the tests
 check every sample the object outputs, not the driver). Changing the sample rate in Audio Status is