Archived
will
d3495476f4
El SDK Release / build-and-release (push) Failing after 13m0s
The generated C, amalgams, vendored runtime pins, and compiled binaries from the Claude Code era are removed from the worktree. The El sources survive; this tree is now source-only for the first-principles rebuild. Per Principal direction 2026-08-19.
169 lines
8.2 KiB
Markdown
169 lines
8.2 KiB
Markdown
# CLAIMS — cycle 16, tokens carry no position
|
||
|
||
Source doc: `docs/v1/experiments/cycles/16-tokens-carry-no-position.md`
|
||
|
||
## Pinned commits
|
||
|
||
| role | sha | subject |
|
||
|---|---|---|
|
||
| parent / before-state | `1086ac965` | record the module answer: the partition is a path, not a neighbourhood |
|
||
| experiment (system under test) | `6c975b1d5019446c29d84d99b9b1c0ec512be8a6` | thread provenance through resolve_imports |
|
||
| land (merge) | `cb7289f06510c002b6ab7c5ca8c2bd182d8216bc` | thread provenance end to end: a diagnostic can finally name a place |
|
||
|
||
## Verdicts
|
||
|
||
| # | claim | artifact | commit | verdict |
|
||
|---|---|---|---|---|
|
||
| A1 | tokens had **no position at all** — a token was a flat `(kind, value)` pair | `0001-pre-token-is-flat-pair.out` | `1086ac9` | REPRODUCED (exact) |
|
||
| A2 | no diagnostic in El could name a place — every error named a symbol, never a line | `0001` (`tok_line` absent from all five compiler sources) | `1086ac9` | REPRODUCED |
|
||
| A3 | P1: **15 stride sites**, encapsulated in `tok_kind`/`tok_value` | `0014`, `0015-P1-stride-sites-all-files-clean-tree.out` | `1086ac9`→`6c975b1` | REPRODUCED (see note) |
|
||
| A4 | P2: adding a line field is mechanical; P3: the lexer must count newlines — stride **2 → 3** | `0014`, `0015` | `6c975b1` | REPRODUCED |
|
||
| A5 | P4: `resolve_imports` can record per-file line ranges (`<file> spans <start> <end>`) | `0009-P5-message-names-both-files.out` | `6c975b1` | REPRODUCED (via effect) |
|
||
| A6 | P5: the message can then name **both files** | `0009` | `6c975b1` | REPRODUCED (exact shape) |
|
||
| A7 | P6: token memory grows — **25.0 → 33.9 MB (+36%)** | `0005`, `0006`, `0007-P6-rss-five-runs-each.out`, `0008-P6-growth-arithmetic.out` | `1086ac9`/`6c975b1` | **DIVERGED** (endpoints reproduce; the percentage does not) |
|
||
| A8 | 105/105 native | `0011-post-native-suite.out` | `6c975b1` | REPRODUCED (exact) |
|
||
| A9 | 37/37 integration | `0012-post-integration-harnesses.out` | `6c975b1` | REPRODUCED (exact) |
|
||
| A10 | fixpoint ok | `0010-post-fixpoint.out` | `6c975b1` | REPRODUCED |
|
||
| A11 | compiler self-checks clean | `0013-post-compiler-self-check.out` | `6c975b1` | REPRODUCED |
|
||
| A12 | defect 1: `interp_tokens_append_all` walks the token list with its own copy of the stride; gen1 built, gen2 emitted corrupt C | — | — | **NOT-CAPTURED** |
|
||
| A13 | defect 2: `tok_count` in `test_compiler.el` carried the stride too, having escaped a compiler-scoped search | `0015` | `1086ac9` | REPRODUCED (corroborated — see note) |
|
||
| A14 | defect 3: nested `resolve_imports` accumulated spans into shared state | — | — | **NOT-CAPTURED** |
|
||
| A15 | defect 4: the first working version reported `b.el:3` — the COMBINED line against a file with no line 3 | `0009` (shows the corrected behaviour only) | `6c975b1` | **NOT-CAPTURED** |
|
||
| A16 | both compilers build | `0003`, `0004` | `1086ac9`, `6c975b1` | REPRODUCED |
|
||
|
||
## A1 / A4 — the token shape
|
||
|
||
`tok_append` at `1086ac9` (`0001`):
|
||
|
||
```el
|
||
fn tok_append(tokens: [Any], kind: String, value: String) -> [Any] {
|
||
let tokens = native_list_append(tokens, kind)
|
||
native_list_append(tokens, value)
|
||
}
|
||
```
|
||
|
||
Two slots. `tok_line` does not appear anywhere in `lexer.el`, `parser.el`,
|
||
`codegen.el`, `codegen-js.el` or `compiler.el` — count 0 in all five. At
|
||
`6c975b1` a third slot is appended from `state_get("__lex_line")` and
|
||
`tok_line(tokens, pos)` is defined at `parser.el:41`.
|
||
|
||
## A3 / A13 — the stride count, and the escapee
|
||
|
||
`0015`, run against `1086ac9`, finds **16** sites that hardcode the stride:
|
||
|
||
```
|
||
codegen.el 3 sites (3633, 3807, 4018)
|
||
parser.el 12 sites (20, 21, 31, 38, 42, 49, 1233, 2161, 2204, 2256, 2422, 2564)
|
||
tests/native/test_compiler.el 1 site (21)
|
||
```
|
||
|
||
Compiler sources alone: **3 + 12 = 15** — exactly P1's figure. The sixteenth is
|
||
`tok_count` in `test_compiler.el`, which is precisely what defect 2 in the
|
||
record says: *"tok_count in test_compiler.el carried the stride too. I had
|
||
scoped the search to compiler sources and it had escaped into the tests."*
|
||
|
||
So P1's 15 is the compiler-source count, it is correct as such, and the
|
||
independent count corroborates the recorded defect rather than contradicting the
|
||
prediction. After the change `parser.el` exposes three accessors
|
||
(`tok_kind:25`, `tok_line:41`, `tok_value:48`) with **187** call sites (`0014`).
|
||
|
||
## A6 — the diagnostic, exactly as recorded
|
||
|
||
Fixtures `fixtures/a.el`, `fixtures/b.el`, `fixtures/m.el`. At `6c975b1`
|
||
(`0009`, checker exit 1):
|
||
|
||
```
|
||
duplicate definition: 'helper' is defined 2 times — El has no namespacing, so imported modules share one global scope
|
||
…/mod16/a.el:1
|
||
…/mod16/b.el:1
|
||
```
|
||
|
||
Same shape as the record's
|
||
|
||
```
|
||
/tmp/modtest/a.el:1
|
||
/tmp/modtest/b.el:1
|
||
```
|
||
|
||
(the directory differs because the fixture was re-created in this re-run's
|
||
scratch space; `/tmp/modtest` from the original session no longer exists). Both
|
||
files are named, and each line number is the line **within that file** — which
|
||
is defect 4's repair: the combined-source line is no longer reported against a
|
||
file that has no such line.
|
||
|
||
## A7 — the divergence, in full
|
||
|
||
Peak RSS of the same compile (`elc-cli.el`), five runs each, `/usr/bin/time -l`
|
||
(`0007`):
|
||
|
||
```
|
||
pre run1 25018368 23.86 MiB 25.02 MB post run1 35536896 33.89 MiB 35.54 MB
|
||
pre run2 25001984 23.84 MiB 25.00 MB post run2 33980416 32.41 MiB 33.98 MB
|
||
pre run3 25067520 23.91 MiB 25.07 MB post run3 35061760 33.44 MiB 35.06 MB
|
||
pre run4 24887296 23.73 MiB 24.89 MB post run4 34947072 33.33 MiB 34.95 MB
|
||
pre run5 25067520 23.91 MiB 25.07 MB post run5 34848768 33.23 MiB 34.85 MB
|
||
```
|
||
|
||
**Both endpoint numbers in the record reproduce — but not in the same unit.**
|
||
|
||
* `25.0` is the **pre** figure read in decimal MB (25.00–25.07 MB; the same
|
||
bytes are 23.84 MiB).
|
||
* `33.9` is the **post** figure read in **MiB** (33.89 MiB on run 1; the same
|
||
bytes are 35.54 MB).
|
||
|
||
Computed in consistent units (`0008`):
|
||
|
||
```
|
||
pre mean 25008537 bytes = 23.84 MiB = 25.00 MB
|
||
post mean 34874982 bytes = 33.25 MiB = 34.87 MB
|
||
growth, consistent units = 39.45 %
|
||
claimed 25.0 -> 33.9 implies 35.60 %
|
||
```
|
||
|
||
So the true growth is **≈ +39.5 %**, not +36 %. The +36 % figure is the
|
||
arithmetic of one decimal-MB reading against one MiB reading. P6's prediction
|
||
("token memory grows") is REPRODUCED and the direction is unaffected; the
|
||
magnitude as printed is DIVERGED, and understates the cost.
|
||
|
||
A note on what is being measured: `/usr/bin/time -l` reports process peak
|
||
resident set, not the token list in isolation. The record does not state the
|
||
method used for its own figures. Since both of its endpoints fall inside the
|
||
peak-RSS distribution measured here, peak RSS is the most likely original
|
||
instrument; but the +36 % is wrong under any single consistent unit, which is
|
||
what the divergence rests on. A stride change from 2 to 3 slots implies +50 % on
|
||
the token list alone, and +39.5 % on whole-process peak is consistent with the
|
||
token list being the dominant but not sole allocation.
|
||
|
||
## Suite results at `6c975b1`
|
||
|
||
```
|
||
105 tests, 105 passed, 0 failed, 253 assertions (0011)
|
||
FIXPOINT-OK gen2 == gen3 (0010)
|
||
definitions: clean exit=0 (compiler's own source) (0013)
|
||
|
||
arity_query.sh exit=0 assertions=5
|
||
capability_query.sh exit=0 assertions=5
|
||
definitions_query.sh exit=0 assertions=6
|
||
prohibition_query.sh exit=0 assertions=4
|
||
seam_binding.sh exit=0 assertions=7
|
||
temporal_query.sh exit=0 assertions=6
|
||
temporal_signatures.sh exit=0 assertions=4
|
||
TOTAL: 37 (0012)
|
||
```
|
||
|
||
105/105 and 37/37 — both exact.
|
||
|
||
## Provenance notes
|
||
|
||
`0002` carries a DIRTY-TREE flag (`git show` measurement run from the shared main
|
||
checkout while other agents were committing to it). It was re-run from a clean
|
||
worktree as `0014`, and extended as `0015`. Cite the `-clean-tree` artifacts.
|
||
Everything else is from `/tmp/rerun-v1b-16-pre` (`1086ac9`) and
|
||
`/tmp/rerun-v1b-16-post` (`6c975b1`).
|
||
|
||
The four defects the record lists (A12–A15) were found and fixed *within* the
|
||
development of `6c975b1`; no intermediate states were committed, so no artifact
|
||
of them exists and none was manufactured. A13 is graded REPRODUCED only because
|
||
the escapee it names is independently visible in the stride census at the parent
|
||
commit.
|