plans: Phase 9 — make .apl source files run as-written

Goal: existing lib/apl/tests/programs/*.apl execute through
apl-run unchanged.  Sub-tasks: compress-as-fn (mask/arr),
inline assignment, ? random, apl-run-file, end-to-end .apl
tests, glyph audit.

plans/apl-on-sx.md

@@ -177,6 +177,42 @@ programs run from source, and starts pushing on performance.
   300 s timeout). Target: profile the inner loop, eliminate quadratic
   list-append, restore the `queens(8)` test.
 
+### Phase 9 — make `.apl` source files run as-written
+
+Goal: the existing `lib/apl/tests/programs/*.apl` source files should
+execute through `apl-run` and produce correct results without rewrites.
+Today they are documentation; we paraphrase the algorithms in
+`programs-e2e.sx`.  Phase 9 closes that gap.
+
+- [ ] **Compress as a dyadic function** — `mask / arr` between two values
+  is the classic compress (select where mask≠0).  Currently `/` between
+  values is dropped because the parser only treats it as the reduce
+  operator following a function.  Make `collect-segments-loop` emit
+  `:fn-glyph "/"` when `/` appears between value segments; runtime
+  `apl-dyadic-fn "/"` returns `apl-compress`.  Same for `⌿`
+  (first-axis compress).
+- [ ] **Inline assignment** — `⍵ ← ⍳⍵` mid-expression.  Parser currently
+  only handles `:assign` at the start of a statement.  Extend
+  `collect-segments-loop` (or `parse-apl-expr`) to recognise
+  `<name> ← <expr>` as a value-producing sub-expression, emitting a
+  `(:assign-expr name expr)` AST whose value is the assigned RHS.
+  Required by the primes idiom `(2=+⌿0=⍵∘.|⍵)/⍵←⍳⍵`.
+- [ ] **`?` (random / roll)** — monadic `?N` returns a random integer
+  in 1..N.  Used by quicksort.apl for pivot selection.  Add `apl-roll`
+  (deterministic seed for tests) + glyph wiring.
+- [ ] **`apl-run-file path → array`** — read the file from disk, strip
+  the `⍝` comments (already handled by tokenizer), and run.  Needs an
+  IO primitive on the SX side.  Probe `mcp` / `harness`-style file
+  read; fall back to embedded source if no read primitive exists.
+- [ ] **End-to-end .apl tests** — once the above land, add tests that
+  run `lib/apl/tests/programs/*.apl` *as written* and assert results.
+  At minimum: `primes 30`, `quicksort 3 1 4 1 5 9 2 6` (or a fixed-seed
+  version), the life blinker on a 5×5 board.
+- [ ] **Audit silently-skipped glyphs** — sweep `apl-glyph-set` and
+  `apl-parse-fn-glyphs` against the runtime's `apl-monadic-fn` and
+  `apl-dyadic-fn` cond chains to find any that the runtime supports
+  but the parser doesn't see.
+
 ## SX primitive baseline
 
 Use vectors for arrays; numeric tower + rationals for numbers; ADTs for tagged data;
@@ -191,6 +227,7 @@ data; format for string templating.
 
 _Newest first._
 
+- 2026-05-07: Phase 9 added — make .apl source files run as-written (compress as dyadic /, inline assignment, ? random, apl-run-file, glyph audit, source-as-tests)
 - 2026-05-07: Phase 8 step 6 — perf: swapped (append acc xs) → (append xs acc) in apl-permutations to make permutation generation linear instead of quadratic; q(7) 32s→12s; q(8)=92 test restored within 300s timeout; **Phase 8 complete, all unchecked items ticked**; 497/497
 - 2026-05-07: Phase 8 step 5 — train/fork notation. Parser :lparen detects all-fn inner segments → emits :train AST; resolver covers 2-atop & 3-fork for both monadic and dyadic. `(+/÷≢) 1..5 → 3` (mean), `(- ⌊) 5 → -5` (atop), `2(+×-)5 → -21` (dyadic fork), `(⌈/-⌊/) → 8` (range); +6 tests; 496/496
 - 2026-05-07: Phase 8 step 4 — programs-e2e.sx runs classic-algorithm shapes through full pipeline (factorial via ∇, triangulars, sum-of-squares, divisor-counts, prime-mask, named-fn composition, dyadic max-of-two, Newton step); also added ⌿ + ⍀ to glyph sets (were silently skipped); +15 tests; 490/490