[tap]

What this changes

Docs only, four commits:

1. The record brought up to 1.0.2. CLAUDE.md's rule about the reserved-name guard describes 1.0.2, not 1.0.1, and docs/AUDIT-TAP-PYTHON-PLAN.md gains a dated addendum on what the two releases since the audit changed in its findings.
2. docs/TAP-PYTHON-PLAN.md revised for every audit finding, with the production plan's Phase 9 and sequencing following. The plan ends with a revision record naming what each finding changed.
3. After the maintainer's review: documentation in Max's own system (9.4), a book (9.7), and the test coverage gaps closed in 9.1, 9.3 and 9.5.
4. The three open decisions, decided by the maintainer: the release that carries tap.python is 2.0.0 (the list[…] parameter is a contract change, the CHANGELOG's rule applied as written); a str return outputs as the spike finds composes best; the vignette is the short form, written separately.

Why

The audit found two blockers in the plan as first written. The revision settles them:

• One binary, two Max classes (D11). Two externals would each compile in their own copy of the header-only core and start the interpreter; the second aborts Max (reproduced with two shared objects in one process), and 1.0.1's Windows failure showed a function's address is not its identity across a DLL boundary. The maintainer chose one binary on 2026-10-09.
• The second class is a plain SDK class, not a min class. min keeps one file-scope this_class per translation unit (c74_min_api.h:325) and returns early from a second wrap_as_max_external(). A second min class in another translation unit would work only by accident of internal linkage, and break the one-definition rule the moment a test instantiated its templates elsewhere. A plain class is what the shared glue already speaks, and it is what the audit's outlet finding needed anyway. Max finds it through the package's init/ objectfile mapping, with a stub external as the fallback; the spike settles which. Its reference page is written by hand, since there is no min generator for it.
• A spike in Max, on a Mac and on Windows, before any code the rest depends on (9.0). Three releases in a row shipped what only a host platform could show, and the third was Windows-only. The spike's questions are listed in the plan.

The major and minor findings are folded in as the revision record says: raw outlets created dumpout-first with the dumpout in the obex; a recursive per-object lock across output and port changes; a class-level anything forwarder that fixes the dispatch order itself; Scheduler in Audio Interrupt detected and stated as a limit; announce-once per object kind; the tuple rule gated on bind_audio; the list[…] parameter recorded as a change with the old behavior pinned first.

The review added what the plan lacked. Documentation was reference pages and help files only; 9.4 now has a vignette, three tutorials with patchers, an Extras overview named as the landing patcher, help tabs for both objects, and CI checks for what is written by hand, with 9.0 finding where Max reads the files from. 9.7 is the family's mdBook after 2.0.0, with example code and performance tables pulled from what ships. Test coverage had four gaps: no sanitizer rows for the Max glue, where the new lock lives; a performance sentence with no measurement; no soak for the new object; and the hand-made pages, help patchers and init/ file unchecked. All four are in the plan now, plus the output table as a data-driven test over its edges, core tests for the new examples, and an injected audio-thread predicate so the Scheduler-in-Audio-Interrupt notice is testable without Max. Windows stays by hand and is written down as a limit.

Verification

Prose only. Every claim about min and the SDK was checked against the headers in the submodule (this_class, wrap_as_max_external, deduce_maxclassname, doc_update, object_obex_store, outlet_insert_after, systhread_isaudiothread) and the mock kernel's stubs; the claims about the releases against the commits on main; the claims about Max's package folders against Cycling '74's current package documentation. The plan's section links and its references to the audit's finding ids were checked by script. What only Max can settle is listed under 9.0, not asserted.

Notes for the reviewer

• Contract change. The list[…] parameter changes what a tap.python~ method with such a hint receives (the empty symbol today). Nothing in this PR changes code; the plan records that the release carrying it is 2.0.0.
• Nothing in the plan is built. 9.0 is the first PR.

🤖 Generated with Claude Code

https://claude.ai/code/session_01WnyasLjC6poziMpwguw4in