1 files changed, 81 insertions, 6 deletions

diff --git a/docs/chapter-backports.md b/docs/chapter-backports.md
index c5a7841d..159bc9e6 100644
--- a/docs/chapter-backports.md
+++ b/docs/chapter-backports.md
@@ -69,7 +69,7 @@ Chapter 25 suite, including `make -j 4 tests`.
 
 | Group | Eventual home | Why not in the small queue yet |
 |---|---|---|
-| Register allocation and spilling | 20 onward | Chapter 20 correction/support pass complete locally; review before Chapter 21. Staged quality work remains below. |
+| Register allocation and spilling | 20 onward | Chapters 20-21 correction/support passes complete locally; review Chapter 21 statistics before Chapter 22. Staged quality work remains below. |
 | Conditional Store and array Load control | Memory/arrays chapters | Reproduce under the earlier alias model before extracting fixes from the new memory implementation. |
 | SCCP dependencies, function revival, reachability | 24; some foundations may fit 18 | Separate old-IR corrections from new Guard/Escape/BulkMemPhi and external-caller machinery. |
 | TypeScalar, numeric modes, guards, symbolic fields, open/forward types | Revisit earlier homes later | A connected incomplete-types architecture, including phase ordering and errors. |
@@ -78,8 +78,8 @@ Chapter 25 suite, including `make -j 4 tests`.
 
 ## Register allocation: correctness first, staged improvements
 
-Review on 2026-09-20. The staged plan follows; Chapter 20
-has since been implemented and validated as recorded below. Later snapshots
+Review on 2026-09-20. The staged plan follows; Chapters 20-21
+have since been implemented and validated as recorded below. Later snapshots
 remain pending. Compared RegAlloc, BuildLRG, IFG,
 LRG, Coalesce, RegMask, and split support across 20-25, with the original fix
 commits. Allocation starts in 20; 19 has instruction selection/register masks
@@ -153,16 +153,16 @@ to add persistent deferral state to the first allocator chapter.
   of the heuristic's exact spill count.
 - `splitBypass` originally scanned from `j-1` with `idx++` throughout 20-25,
   reaching its own destination and rejecting nonadjacent bypasses. Corrected
-  in 20 with intervening kill-mask checks and a regression; 21-25 remain pending.
+  in 20-21 with intervening kill-mask checks and a regression; 22-25 remain pending.
   Pre-color copy reuse also checks fixed-register definitions before they have
   an assigned `_reg`. Chapter 25 is not a complete correctness reference.
 
 Suggested execution order: support/measurement, small mask/LRG correctness
 fixes, constrained-register and self-conflict regressions/fixes, then one
 quality technique at a time. Run each affected snapshot's full suite at each
-accepted boundary. Chapter 20 has now been implemented; do not proceed to 21 until Cliff reviews it.
+accepted boundary. Chapters 20-21 have now been implemented; stop for Cliff's review before 22.
 Review each chapter's README along the way. Chapter 21's encoding discussion
-should be shortened when that chapter is reached. End each README with its
+has been shortened, with the bit-level notes retained as a separate reference. End each README with its
 RegAlloc improvement, measured cohort table, and commentary; the Chapter 25
 table must have rows for cohorts 20-25 using the Chapter 25 compiler.
 
@@ -206,6 +206,81 @@ Historical results below are dated evidence, not a substitute for a fresh
 baseline. Logs live in ignored build directories and may no longer exist.
 Reusable implementation lessons are in `skills/chapter25-codex-notes.md`.
 
+### Inlined return typing and ARM emulator: corrected, 2026-09-20
+
+The ARM seed-9 String failure was an optimizer bug: inlining can delete a FunNode
+while its Return still carries live control, memory, and data. ReturnNode.compute
+mistook the deleted entry for an unreachable return and supplied Top to its
+caller. Chapter 20 introduced direct use of the linked Return's type, exposing
+this stale guard. Removed it in 20-21, matching the existing 22-25 implementation.
+Chapter 19 does not read that linked return type and passes the reduced probe.
+
+Reduced source: `struct S { int x; }; val f={ S s -> s.x=g(); };
+val g={ -> 123; }; S !s=new S; return f(s);`. The old 20 compiler returns null
+instead of 123 at seed 58; 21 fails at seed 8. The regression checks interpreted
+results across seeds 0-63 in Chapter20Test in every snapshot 20-25.
+
+Execution then exposed missing register-register SUB decoding in EvalArm64.
+Added the unshifted SUB form already emitted by Simple in 21-25. The matching
+Chapter21Test checks positive/negative results, 64-bit overflow, and preserved
+flags; it traps with the old emulator. The original String source at seed 9 now
+encodes and executes on both ARM and RISC-V, producing the independently computed
+hash `-2449306563677080489`. All 1,170 encoding/register checks pass over the frozen
+13-program cohort, three targets, and seeds 0-29.
+
+Fresh baselines passed before edits. Full suites after correction: 20: 376+1;
+21: 401+1; 22: 399+1; 23: 418+1; 24: 441+1; 25's six groups:
+8, 378, 34, 18, 1, 12. Spill totals are unchanged: Chapter 20 remains 236 splits /
+355 weighted; Chapter 21's two cohorts remain 401/653 and 483/1,141. These regressions
+are excluded from spill measurements. Logs/reducers are in each affected chapter's
+ignored `build/arm-top`; the earlier chapter-by-chapter allocator review gate remains.
+
+### Chapter 21 allocator: ready for review, 2026-09-20
+
+Carries the Chapter 20 legality/progress/support fixes into 21 while retaining
+native frame/ABI handling and conservative coalescing. Stronger color bias,
+clone/callee-save ranking, and popular-use grouping are deferred to the planned
+later chapters. The existing loop-entry-tail treatment remains; removing it
+increased weighted counts in a controlled trial. Coalescing's mask, interference,
+capacity rollback, and adjacency-remapping checks pass. Five inherited regression
+cases fail against the original compiler and pass now: mask iteration/empty masks,
+resultless kills, incompatible clone/use classes (round limit), commutative Phi
+uses, and copy cleanup across clobbers. The other two inherited cases already pass.
+
+`make -j 4 tests`: original 385 + 1; corrected 399 + 1 at the allocator boundary (401 + 1 after the return/emulator
+corrections above), after forced rebuilds,
+including native x86/Win64 and RISC-V/ARM emulator checks. All 1,170 allocation
+and register-constraint checks pass for the 13 frozen Chapter 20 inputs, three
+CPUs, and seeds 0-29. The subsequent encoding sweep also passes after the return-typing correction
+recorded above; the original ARM seed-9 failure is resolved.
+Native harnesses now retain full process exit codes, and ARM BrainFuck checks its
+own result pointer. `make spill-stats` passes all 45 selected tests (43 at the allocator boundary) and records
+91 compilations; spill-golden mismatches are deferred during reporting so they do
+not suppress later targets/runtime checks, but still make the command fail.
+
+| Cohort, all using Chapter 21 | Compilations | Original splits | Current splits | Original weighted | Current weighted |
+|---|---:|---:|---:|---:|---:|
+| Chapter 20, three SystemV targets | 39 | 340 | 401 | 445 | 653 |
+| Chapter 21, fixed Windows native/emulator suite | 52 | 454 | 483 | 1,175 | 1,141 |
+| Total | 91 | 794 | 884 | 1,620 | 1,794 |
+
+Same corrected compiler without coalescing: 1,074 splits / 1,921 weighted; with
+coalescing: 884 / 1,794, saving 190 / 127. Relative to the original combined
+heuristics, staging costs 174 weighted moves (+10.7%). RISC-V BrainFuck is the
+largest regression; sieve improves on all three targets. Do not claim an overall
+quality improvement over the original snapshot. Chapter-to-chapter raw totals
+also reflect changed lowering and native ABI obligations, not just coalescing.
+
+Frozen Chapter 20 sources (including original BrainFuck/MergeSort) now reside in
+Chapter20Test. The four revised source/ABI cases formerly there are preserved in
+Chapter21AllocTest and counted in 21; its native BrainFuck/MergeSort variants also
+belong to 21. No diagnostic graphs count toward quality totals. The README now
+summarizes encoding/ELF, links the retained encoding reference, and ends with
+coalescing, both cohort rows, and the controlled comparison. Logs and isolated
+baseline/ablation sources are under `chapter21/build/regalloc-review` (ignored).
+The allocator changes remain confined to Chapter 21; the subsequent return/emulator
+correction is recorded separately above. Chapter 22 allocator work awaits review.
+
 ### Chapter 20 allocator: ready for review, 2026-09-20
 
 Implemented only in 20, per Cliff's chapter-by-chapter review gate. Corrected