summaryrefslogtreecommitdiff
path: root/skills/chapter25-codex-notes.md
diff options
context:
space:
mode:
Diffstat (limited to 'skills/chapter25-codex-notes.md')
-rw-r--r--skills/chapter25-codex-notes.md25
1 files changed, 25 insertions, 0 deletions
diff --git a/skills/chapter25-codex-notes.md b/skills/chapter25-codex-notes.md
index e7362c3b..b32fdd15 100644
--- a/skills/chapter25-codex-notes.md
+++ b/skills/chapter25-codex-notes.md
@@ -35,6 +35,31 @@ callers promise no initialized fields. No new passes or backward demand walks.
## Chapter 32: restart guidance (2026-11-15)
+Use `make -j8 tests` as the normal validation/timing command going forward.
+The November 16 sweep found both `-j4` and `-j16` slower on this Windows host;
+retain `ActiveProcessorCount=2`. Keep lattice checks in their own JVMs: combining
+them after compiler-heavy tests can exhaust the existing 16-bit type IDs.
+The full lattice is required even in fast checks. `make -j8 sanity` keeps its
+four partitions plus chapter/backend and focused compiler regressions; full
+`make -j8 tests` retains library/application/CLI integration. The lattice now
+caches pairs and checks all six intermediate results per unordered triple;
+do not reduce it to one associativity equality per unordered triple, which
+would assume commutativity for intermediate types outside the gathered set.
+
+On 2026-11-16, Cliff requested backporting the shared class receiver to chapter
+28a, imported-summary handling to 28b, and both through the later snapshots.
+The chapter28a README now owns that explanation. Preserve 28a/28b's monolithic
+parser and Scope-based instance construction when adapting this fix; do not
+copy chapter29's parser or construction machinery backward. Native fixtures
+must also respect the older chapters' 64-bit `int` ABI.
+
+The Phi factoring fix starts in chapter5: match `x*2` as `x+x` without expanding
+the graph. Chapter14 also recognizes `x<<1`; chapter19 and later use that shift
+as doubling's strength-reduced representation. Keep each snapshot's existing
+factoring guards, type checks, and dependency direction. The chapter5 README
+teaches the rule; chapter32 retains the investigation and measurements. No
+allocator scoring experiments were adopted or backported.
+
Continue in `chapter32/` on branch `ch32`. Chapter 31 is complete as the
graphics, games, and library snapshot; keep its compiler before the speedups.
The chapter32 README teaches the current performance work: parallel test JVMs