[timkpaine]

The first repository to flip, read against the principles in FLIP_READINESS.md by an adversarial review. Every accepted finding lands here with a test.

Behaviour

• plan_split("adaptive") refuses with PostureError saying where adaptive is planned, instead of NotImplementedError from a public name. The posture stays in the taxonomy because every consumer speaks it.
• WorkSplit, QueryProfile, NodeCost and LinkRate gain check(), called by the validator and by every use site (execute_split, estimate, is_stale_for, is_stale, both cache writes). model_copy(update=...) skips validation and these models are copied; tests pin that an invalid copy is refused where it is used. extra="forbid" on all of them. SplitExecution validates its own figures.
• RowRange.slice and RowBatch.slice require 0 <= start <= stop <= len; RowBatch refuses a negative start. Python slicing would clamp or wrap these silently.
• profile_from_spans refuses a reversed or negative span.
• ProfileCache expands ~, percent-encodes every key component and refuses any entry that is not a file directly under the root. Entries written under the old naming are not found and get re-measured.

Claims

• The posture docstring's 27.6 / 18.6 ms comparison is restated with the 11.9 ms split from the 2026-08-12 ledger entry; its "40%" sentence was wrong. The 11.2 / 14.6 ms plan-time figures had no ledger row and are gone. "~1.6x between hosts" is 1.41x (2026-08-16); "13 ms host" is 12.3 ms. Analytic examples are labelled analytic.
• README: "declines to bother", "enforces it" and "allocates nothing per call" overclaimed and are corrected; the distinct-count statement is about exact counts; every code block now runs as written and test_readme.py runs them in order; an install section says how to install from the repository.

Quality gates

• ARG enabled (tests exempt, per the review rule) and the two production hits fixed; check-types is part of make checks and its 21 diagnostics are fixed; classifiers trimmed to the CPython 3.12 CI tests; orphan coverage omit removed; the vacuous test_all asserts the exported surface.
• Wiki: Python 3.12 to match the metadata, no conda-forge claim, install from the repository, headings no longer named after steps.

Dispositions of the findings not changed here (posture taxonomy, host as measurement data, template-managed workflow and Makefile items) are recorded in FLIP_READINESS.md.

[github-actions]

Test Results

115 tests  +16   115 ✅ +16   1s ⏱️ ±0s
  1 suites ± 0     0 💤 ± 0 
  1 files   ± 0     0 ❌ ± 0 

Results for commit a221e68. ± Comparison against base commit a2d3ea8.

This pull request removes 2 and adds 18 tests. Note that renamed tests count towards both.

dau_scheduler.tests.test_all ‑ test_all
dau_scheduler.tests.test_posture ‑ test_every_posture_is_reachable_or_explicitly_unimplemented

dau_scheduler.tests.test_all ‑ test_the_documented_surface_is_exported
dau_scheduler.tests.test_posture ‑ test_every_posture_is_reachable_or_refused_by_name
dau_scheduler.tests.test_profile ‑ test_a_copied_link_rate_is_checked_where_it_is_used
dau_scheduler.tests.test_profile ‑ test_a_copied_profile_is_checked_where_it_is_used
dau_scheduler.tests.test_profile ‑ test_a_reversed_or_negative_span_is_a_reporting_fault_not_a_cheap_node[spans0]
dau_scheduler.tests.test_profile ‑ test_a_reversed_or_negative_span_is_a_reporting_fault_not_a_cheap_node[spans1]
dau_scheduler.tests.test_profile ‑ test_cache_entries_stay_directly_under_the_root
dau_scheduler.tests.test_profile ‑ test_models_refuse_fields_they_do_not_declare
dau_scheduler.tests.test_profile ‑ test_the_cache_root_expands_a_home_directory
dau_scheduler.tests.test_readme ‑ test_the_readme_examples_run_in_order
…