PR #107 (instrument skip rework) shipped a few days after the original
arc closed. Captured three new datapoints worth recording:

- Mid-design pivot worked: user pushback on the is_configured arg
  forced the design through pre-#88 history and confirmed the
  classmethod-with-arg pattern was correct.
- Defensive workaround was wrong design: my _get_logger() fresh-per-
  call structlog proxy made tests pass but the user's pivot to stdlib
  logging + public build_summary() was the actual right answer. caplog
  (which the original plan flagged but the subagent ignored) was the
  correct test mechanism.
- LSP violations on classmethod parameter overrides are an emergent
  pattern: 3 framework instrument overrides needed
  # ty: ignore[invalid-method-override] because they narrow
  bootstrap_config covariantly. Existing field-narrowing pattern is
  accepted; method-param narrowing now needs the explicit ignore.

Added action items #8 (workaround-vs-framework-choice heuristic), #9
(cap subagent dispatch scope; the ~60-minute connection drop orphaned
work), and #10 (document the LSP override pattern in CLAUDE.md).

Reaffirms the original retro's closing observation: subagent loop
reliably produces a green-tests implementation of the spec, but the
spec is rarely the right design. Design emerges during review
iteration. PR #107 needed 5 user follow-up commits after the
mechanical migration landed.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>