The census first, because it decides how large the mechanical edit is
(framing §7). Every occurrence was listed with its enclosing context and
read; a grep for the bare name over-counts, which is how revision 1
reported 18 by grepping `Editor::new` — a pattern that does not match the
real constructor.
in-process 342 calls in 66 of 97 files
(330 of 334 `EditorState::new()` occurrences; 4 are prose)
(12 of 14 `EditorState::open(` occurrences; 2 are strings)
spawned 14 real `pmacs` spawns in 8 files
(of 36 `CARGO_BIN_EXE_pmacs` hits, 18 are the fake-LSP and
fake-MCP siblings and 4 are path derivations for
`pmacs-gpu`, not spawns)
mixed 5 files are both, so sites — not files — are the unit
The full census, with per-site attribution, is the module doc of
`tests/ambient_isolation_acceptance.rs`.
Four things it pins:
* Isolated construction still finishes initialization, asserted twice —
the flag, and the behaviour it gates (`pmacs.attach` must refuse).
Falsified by wrapping the config block in `if roots.is_ambient()`;
`m8_2_acceptance` does NOT catch that, because reopening an already-open
init phase is a no-op.
* The writes land in the redirected data root — content produced, not an
invariant preserved. A "the real root did not change" check would pass
vacuously wherever it already holds identical bytes, since
`write_if_changed` is content-gated.
* Bet 3, in two children with opposite jobs. The positive control proves
the hostile environment IS hostile (an ambient editor loads its
`init.lua` and writes its data root); without it the isolation half
asserts nothing. The isolated child then stays green under the same
environment and leaves its hostile root byte-identical.
* A durable adoption ratchet, not a one-time census: a source scan that
fails when a new ambient constructor appears outside a named allowlist,
plus a check that no allowlist entry has gone dead. Its scanner strips
comments, strings and raw strings, and that stripping has its own pin —
the corpus contains all three shapes, and a grep-shaped answer already
cost this lane a review round.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Lv428Fth9LRtffwJSsqH7T
Post-ship-gate completion of M10.10 (optimistic local edit
application, Path β). The milestone core landed in 45be65b
"M10.10 ship gate"; this commit completes the frontend keystroke
path and absorbs Day-5 corrections.
Completes:
- optimistic::frontend_event_for_keystroke — keystroke orchestrator
(classify_key predicate → mirror-ready check → CrdtOp or Key
fallback per Refinement 4 graceful degradation).
- BufferMirror cursor tracking (active_buffer, cursor_byte_pos via
CursorByte) + char-boundary-aware delete helpers (prev/next_char_len)
so multibyte backspace/delete don't trip loro's mid-codepoint
rejection.
- buffer.rs: crdt_state accessor (was test-only) now production —
daemon's BufferSnapshot export path uses it.
Day-5 corrections:
- packages/manifest.rs: fix pre-existing M8-era proptest generator
that produced ".."-containing entry paths the parser correctly
rejects (segment-structured regex; stale regression seed removed).
Out of M10.10 scope; absorbed so future milestone sweeps see clean
output instead of a known-failing test requiring prose.
- tests: extract inline PTY/daemon helpers to shared tests/common/
module (m5_5, m5_8 now import; no coverage change — m5_5 retains
19 m10_10 tests). tests/common/ added (required for compilation).
Audit history (M10.10-AUDIT.md is gitignored internal-only; this
message is the sole version-controlled record):
M10.10 PASSES within Path β scope (end-of-line optimistic visual
paint; mid-line/delete-forward round-trip; full CRDT-op exchange
across the text-input scope). The initial audit verdict was WRONG —
optimistic-apply was structurally unreachable in the production
binary (build_capabilities advertised crdt_replica: false). Six
post-audit review rounds surfaced 28 findings (F5–F32) beyond the
framing pass's original 4. Six M10-era discipline additions emerged,
each empirically grounded: end-to-end-exercise check (bidirectional
scope), composition-consistency check, verification-milestone premise
check, library-API verification check, forward-pointer-comment
hygiene, methodology-composition check.
Scorecard adopts Option C dual methodology: layer (a) framing-pass
accuracy is 7/8 milestones-not-findings AND 2/8 findings-as-failures
— the 5/8 spread is the density diagnostic (M10.10's defining
characteristic; neither number alone is honest). Layer (c): 7/8 and
6/8 (M10.8 inherited-gap cluster). Budget honesty: 5-day
pre-authorization covered anticipated implementation surprises (K1,
Risk #6 a); Finding 3 was a third surprise absorbed via compression,
not structural slack; the six post-audit rounds were entirely
unbudgeted and are the milestone's dominant cost. M10.10's density
is partly forecastable — it is the only M10 milestone with all three
of: architectural reversal, multi-milestone integration, and
verification depending on incomplete cross-milestone wiring.
Ship-gate clean on clean checkout (cargo clean + rebuild): luajit+crdt
1364/1364, luajit 1211/1211, lua54+crdt 1364/1364, lua54 1211/1211,
m5_5 daemon-e2e 36/36, perf 1MB=1.1ms vs 10ms gate, clippy 0 across
feature combos, fmt clean. One transient flake observed
(async_runtime::supersede_cancels_in_flight_job_within_50ms — timing
test starved under concurrent compile load, non-reproducible in
isolation, known infra pattern, not an M10.10 regression).
Next: M10.11 (two-laptop acceptance) inherits all six discipline
additions; v1.0 ships after M10.11.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Levi Neuwirth