el-retired/docs/v1/experiments/evidence/cycles/07-invocation-is-not-composable/CLAIMS.md

CLAIMS — cycle 07, invocation is not composable

Source doc: docs/v1/experiments/cycles/07-invocation-is-not-composable.md

Pinned commits

Cycle commit (the EXPERIMENT)	bc2f26ddfc35119e242c0a38d09e2b0023136462
Committed	2026-08-17 09:06:05 -0500
Subject	EXPERIMENT: invocation control resolves at runtime
Parent (the "before" state the record describes)	b40754f07b5eff3700a69681397e46949a5c4f41 — "land unconditional wrapper: exit crossings resolve at runtime", 2026-08-17 09:01:55 -0500
Land commit (second record block, no numeric claims)	c04d68f9c… — "land runtime invocation control: only prohibition remains compiled", 2026-08-17 09:06:16 -0500
Re-run method	detached worktree /tmp/rerun-v1-07 at bc2f26d, tree clean
Artifact provenance	every artifact pinned to bc2f26ddfc35, tree clean (see MANIFEST.tsv)
Machine	Apple M4 Pro, Darwin 25.5.0 arm64, Apple clang 21.0.0, load avg ≈5–6 during timing

c04d68f is tree-identical to bc2f26d. git diff --stat bc2f26d c04d68f is empty and codegen.el hashes identically at both (0015). The land commit changes no file; it only re-records the experiment as landed. Every measurement below therefore applies to both.

Verdicts

#	Claim (verbatim from the record)	Artifact	Commit	Verdict
1	"P1 wrap becomes runtime-bindable TRUE body x3 -> 21, never invoked -> 111"	0016-P1-wrap-probe-reconstructed.out (reconstructed probe), 0040-REDO-seam-caller-and-wrap-assertion-search.out	bc2f26d	REPRODUCED (behaviour) / DIVERGED (evidence) — a reconstructed probe does observe 21 and 111. But no test or harness in the tree at bc2f26d asserts either number, and none ever exercises the wrap seam at all. See divergence detail A.
2	"P2 codegen shrinks TRUE 5042 -> 4977"	0009-codegen-linecount-parent-and-commit.out, 0011-codegen-linecount-every-commit-in-range.out	bc2f26d^, bc2f26d	DIVERGED — observed 5043 -> 4976 (−67). Claimed 5042 -> 4977 (−65). Neither endpoint matches; no commit in 5718943^..c04d68f has codegen.el at 5042 or 4977 lines. Direction of the claim (it shrinks) holds. See detail B.
3	"P3 cost 5-10% from an indirect call on every fn TRUE 0.36s -> 0.39s, ~8%"	0036-P3-timing-parent-vs-commit-gen3.out, 0037-P3-timing-two-workloads-n21.out	bc2f26d^ vs bc2f26d	DIVERGED on the ratio; NOT-REPRODUCIBLE on the absolutes. Re-measured overhead is +0.6 % to +2.3 %, not ~8 %. No workload in the tree runs anywhere near 0.36 s. See detail C.
4	"P4 zero-param fns break on the empty struct TRUE — empty struct is a GNU extension, empty init is C23. Fixed with a char field."	0017-P4-zero-param-env-struct-emitted-c.out, 0005/0041-native-compiler-tests*.out	bc2f26d	REPRODUCED — emitted C contains struct __env_noargs { char __e0; }; and struct __env_noargs __env = { 0 };; no empty struct, no empty initialiser; the emitted C passes -std=c11 -pedantic-errors -fsyntax-only. The char __e0 return is at codegen.el:3226 at bc2f26d and absent at the parent. Test zero-param-fn-emits-valid-c passes.
5	"P5 fixpoint holds TRUE"	0038-selfhost-fixpoint-gen2-gen3.out (and 0008-REDO-…)	bc2f26d	REPRODUCED — gen2 == gen3, identical sha256 02c9b3e2…, 14360 lines each. gen1 (from the committed dist/platform/elc) differs, which is expected because this commit changes codegen's output. See measurement defect 1.
6	"PROCESS FAILURE … The build then failed with 'undeclared identifier __thunk_noargs'"	0039-thunk-noargs-and-seam-caller-search.out	range 5718943^..c04d68f	NOT-CAPTURED — the state described was never committed. The literal __thunk_noargs appears in no committed file at any commit in the range; it is a name codegen generates (codegen.el:3356). In the C emitted at bc2f26d the thunk is declared at line 12 and used at line 19, and the file compiles and runs clean. The failure is a real-but-transient working-tree state with no artifact.
7	"Removed: declare_wrap, decorator_wrap, cg_wrap_target, cg_wrap_construct, params_to_call_args, and the wraps_body scanner branch"	0012-removed-symbols-parent-vs-commit.out, 0013-params_to_call_args-orphan-comment.out, 0045-wraps_body-scanner-branch-at-parent.out	bc2f26d^, bc2f26d	DIVERGED (one of six) — five of six are gone from the whole tree. params_to_call_args loses its definition (parent codegen.el:3299) and its only call site (parent codegen.el:3415), but its two-line doc comment survives at bc2f26d codegen.el:3269, describing a function that no longer exists. See detail D.
8	"prohibits_outside is now the ONLY construct kind left at compile time"	0014-construct-kinds-parent-vs-commit.out	bc2f26d	REPRODUCED, with a scope note — prohibits_outside is the only kind name with executable dispatch in codegen.el (str_eq(dkind, "prohibits_outside") at line 4326). injects_at_entry, guards_at_entry, injects_at_exit each survive at exactly one line, all of them // doc comments. wraps_body is gone tree-wide. Note the claim was already true of the other three at the parent — this commit's contribution is removing wraps_body only.
9	(not stated in the doc) native compiler test result at bc2f26d	0005-native-compiler-tests.out, 0041-native-compiler-tests-gen2-binary.out	bc2f26d	REPRODUCED as fact, but the doc records nothing to check it against — 100 tests, 100 passed, 0 failed, 241 assertions. Parent: 99 tests, 99 passed, 0 failed, 239 assertions (0043). The doc gives no test count for this commit; its absence is itself a gap, since the commit deletes a test (declared-wrap-emits-closure-and-convention) and adds two.
10	(not stated in the doc) integration harness result at bc2f26d	0007, 0042, 0043	bc2f26d	REPRODUCED as fact — seam_binding.sh: 7 assertions, 7 passed, 0 failed. At the parent the same harness reports 7 assertions, 6 passed, 0 failed — an arithmetic bug ($((6-FAILS))) that this commit fixes. The commit adds no assertion; its only substantive harness change is appending a dead thrice target.
11	"a construct bound after the build decides how and whether to invoke it" (ISHIKAWA)	0016, 0007/0042	bc2f26d	REPRODUCED — via the reconstructed probe for the wrap phase, and directly by the shipped harness for the entry/exit/refuse phases.
12	(not stated in the doc) bc2f26d silently fixes a harness pass-count bug	0046-seam_binding-tally-history.out	bc2f26d	OBSERVED BUT UNCLAIMED — see detail E.

Counts: 6 REPRODUCED · 4 DIVERGED · 1 NOT-CAPTURED · 1 OBSERVED-BUT-UNCLAIMED · (0 pure NOT-REPRODUCIBLE; claim 3 carries a NOT-REPRODUCIBLE component on its absolute figures).

---

Divergence detail A — P1's 21 and 111 have no evidence in the tree

This is the substantive finding for cycle 07.

The numbers are behaviourally correct — I observed 21 and 111. What is not correct is the implication that the commit demonstrated them.

What bc2f26d actually did to lang/tests/integration/seam_binding.sh (the whole diff, from git diff bc2f26d^ bc2f26d -- lang/tests/integration/):

1. appended a thrice wrap target to $WORK/targets.c, and
2. changed the summary line from $((6-FAILS)) to $((7-FAILS)).

Both are inert with respect to P1:

• thrice is never compiled. targets.c is written at harness line 29 and passed to cc at line 57. The cat >> "$WORK/targets.c" that adds thrice is at line 86 — twenty-nine lines after the binary was already linked. Nothing recompiles it (0040 §5d).
• No wrap-phase binding is ever written. el_seam_load at bc2f26d recognises three phases (entry, exit, wrap); the harness writes entry and exit binding files only. 0040 §5e: no printf '… wrap …' anywhere in the harness. §6: no wrap binding in any .sh/.txt in the tree.
• The strings 21 and 111 do not occur in the harness at all (0040 §5b). Its seven expectations are 7, SEEN work/audited\n7, 42, 7, 7, SEEN work/a\nSEEN work/b\n7, 14 (0040 §5c).
• The commit's own native test comments say the N-times behaviour "lives in tests/integration/seam_binding.sh" — it does not.
• tests/integration/seam_caller.sh does not exist at any commit in 5718943^..c04d68f (0040 §4). It first appears at d9d1571, a v2-era commit, and is unrelated to this cycle. (The parent agent's hint to look for it is answered: not here.)

So the 7→7 pass count reported by the harness at bc2f26d is the same seven assertions the parent already ran; the commit fixed the arithmetic that had been under-reporting them as 6.

The reconstruction (an INPUT, not evidence). evidence/.../probe07.sh, output captured as 0016. It is modelled line-for-line on seam_binding.sh at bc2f26d: the same prog.el (fn work() -> Int { return 7 }), the same thrice target verbatim including its comment, the same EL_CONSTRUCTS file format, and the phase name wrap read from el_seam_load in el_runtime.c at bc2f26d. The one thing not recoverable from the record is the "never invoked" target — the record names no symbol — so never() (returns 111 without calling the body) is invented to match the stated semantics. Observed:

== baseline, no binding (work() returns 7) ==      7
== wrap binding -> thrice (expect 21) ==          21
== wrap binding -> never  (expect 111) ==        111

7 × 3 = 21 corroborates that the record's "body x3" was measured against this same work() -> 7 fixture. 111 is only as faithful as the invented target.

Divergence detail B — P2's line counts

	claimed	observed (wc -l, blob straight out of git)
parent bc2f26d^	5042	5043
commit bc2f26d	4977	4976
delta	−65	−67

Both blobs end in a newline (0009 checks the last byte), so wc -l is exact; there is no trailing-newline ambiguity to absorb the difference. git diff --numstat for codegen.el is 12 79 (+12 / −79 = −67), independently confirming the observed delta.

No commit in 5718943^..c04d68f ever has codegen.el at 5042 or 4977 lines (0011, explicit search: NONE). The full sequence over the range is:

5718943f2  4800   60737b030  4850   4f7568b07  4953   2bed8483f  5084
7d01608a9  5084   1b324a071  5152   35b07bade  5156   82e998273  5152
886626a64  5156   28d19da7f  5095   8bbb750c2  5095   24f7fb514  5095
285166c25  5043   b40754f07  5043   bc2f26ddf  4976   c04d68f9c  4976

Cross-checking the parent agent's note: cycle 03 claims "5094 -> 5044" — observed 24f7fb5 = 5095 → 285166c = 5043. Cycle 06 claims "5157 -> 5096" — observed 35b07ba = 5156, 28d19da = 5095. So every one of these three records is off by one on every endpoint, but not in a consistent direction: cycle 06 is +1 on both ends, while cycles 03 and 07 are −1 on the parent and +1 on the commit. A single systematic counting convention (e.g. counting a missing final newline) would produce a uniform offset; this does not. The line numbers in these records were not produced by a reproducible instrument.

The qualitative claim — "codegen shrinks" — is unambiguously true, by 67 lines.

Divergence detail C — P3's ~8 % cost

The record's absolute figures (0.36s -> 0.39s) are NOT-REPRODUCIBLE: the workload is not named, and nothing in the tree at bc2f26d takes anything like 0.36 s on this machine (0037, tail): elc elc-cli.el 0.067 s, elc --test tests/native/test_compiler.el 0.068 s, elc --test tests/native/test_lexer_scaling.el 0.030 s. The compiled native test binary runs in 5.9 ms. Whatever was timed, it is not identifiable from the record.

The ratio is the load-bearing part, so I re-measured it. Construction (time07b.sh, both scripts stored beside this file):

• The cost under test is paid by code emitted by the new codegen, so both binaries must be gen2 — self-emitted by their own commit's compiler. /tmp/elc07pb = elc self-emitted at bc2f26d^ (wrapper calls the body directly; 0 occurrences of el_seam_wrap in its C); /tmp/elc07b = elc self-emitted at bc2f26d (254 occurrences of el_seam_wrap).
• Both are handed the same input, in the same CWD (bc2f26d's lang/), so the only difference is the calling convention compiled into the binary.
• Runs are interleaved, warmed, n = 21 per arm, two workloads.

Observed (0037):

workload	parent median	commit median	ratio (median)	ratio (min)	ratio (mean)
elc elc-cli.el	0.06482 s	0.06576 s	+1.45 %	+2.31 %	+1.92 %
elc --test tests/native/test_compiler.el	0.06758 s	0.06821 s	+0.93 %	+0.62 %	+1.36 %

An earlier, independent n = 11 run (0036) on the first workload gave +0.64 % median / +0.16 % min. The overhead is real and positive in every arm of every run, but it sits at roughly 1 %, not ~8 % — outside the predicted 5–10 % band, i.e. the prediction as scored would be FALSE on this machine, not TRUE.

Caveats stated plainly, not to rescue the claim: this is an Apple M4 Pro under load avg ≈5–6; el_seam_wrap's fast path is if (_el_seam_n == 0) return body(env); and a wide out-of-order core hides that better than most; and the workload the record actually used is unknown, so I cannot rule out that some other program shows 8 %. No number here is rounded toward the claimed value.

Divergence detail D — params_to_call_args is not fully removed

Of the six things the record says were removed, five vanish from the entire tree between bc2f26d^ and bc2f26d: declare_wrap, decorator_wrap, cg_wrap_target, cg_wrap_construct, and the wraps_body scanner branch (0012; the branch itself is shown at 0045, parent codegen.el:4385, together with the declare_wrap(fname, pending_wrap) call it guarded).

params_to_call_args does not. At the parent it has three occurrences — a doc comment (3297), the definition (3299), and one call site (3415). At bc2f26d the definition and the call site are gone but the comment remains:

bc2f26d:lang/el-compiler/src/codegen.el:3269:// params_to_call_args — "a, b, c" from the param list, for the wrapper's call
bc2f26d:lang/el-compiler/src/codegen.el:3270:// into the body helper.

git grep "^fn params_to_call_args" at bc2f26d → no definition. It is an orphaned two-line comment sitting immediately above fn cg_fn, documenting a function that no longer exists, and it survives unchanged into c04d68f. Harmless, but the record's "Removed:" list overstates by one item, and the residue is exactly the kind of thing the commit message's own "the guard that would have caught it existed and I removed it" paragraph is about.

Detail E — an undocumented fix: the harness pass tally

bc2f26d also fixes a bug its own record never mentions. Independently flagged by the cycle 08 agent and confirmed here (0046):

8bbb750c2  check-calls=6   echo "  6 assertions, $((6-FAILS)) passed, $FAILS failed"
24f7fb514  check-calls=6   echo "  6 assertions, $((6-FAILS)) passed, $FAILS failed"
285166c25  check-calls=7   echo "  7 assertions, $((6-FAILS)) passed, $FAILS failed"   <- seventh check added, tally not bumped
b40754f07  check-calls=7   echo "  7 assertions, $((6-FAILS)) passed, $FAILS failed"
bc2f26ddf  check-calls=7   echo "  7 assertions, $((7-FAILS)) passed, $FAILS failed"   <- fixed here
c04d68f9c  check-calls=7   echo "  7 assertions, $((7-FAILS)) passed, $FAILS failed"

285166c (cycle 03's experiment) added the seventh check and updated the header text to "7 assertions" but left the arithmetic at $((6-FAILS)), so a fully green run printed 7 assertions, 6 passed, 0 failed — confirmed by running the harness at the parent (0043), which prints exactly that with zero failures. bc2f26d changes it to $((7-FAILS)) and the run at this commit prints 7 assertions, 7 passed, 0 failed (0007, 0042).

Worth recording for two reasons. First, a commit doing something its record does not mention is a gap in the record regardless of whether the change is a fix. Second, it is the only substantive thing bc2f26d does to the harness that has any effect — the thrice target it also adds is dead code (detail A) — which sharpens the point that the record's headline P1 result was never wired into the test suite. The harness self-report was under-counting a green run for two cycles before anyone noticed, which is the same class of silent-instrument defect the commit message's PROCESS FAILURE paragraph is about.

---

Measurement defects in this re-run, recorded rather than removed

1. 0006-selfhost-fixpoint.out is defective and retained. It ran the recipe's original fixpoint command, which compares gen1 vs gen2 — gen1 being the output of the committed dist/platform/elc, which is stale at this commit and predates the seam entirely (0 occurrences of el_seam_wrap). Since bc2f26d changes what codegen emits, gen1 ≠ gen2 is expected, and the artifact's differ: char 12626, line 259 is not a finding. The correct check is gen2 == gen3, captured as 0038-selfhost-fixpoint-gen2-gen3.out (and, under a confusing name that used 1-based generation numbering, 0008-REDO-selfhost-fixpoint-gen3-gen4.out — same two files compared, /tmp/g07b.c vs /tmp/g07c.c). This was a recipe defect, not a tree defect; it is left in the manifest because deleting a failed measurement is the failure mode this record exists to control for. Do not cite 0006.
2. 0039-thunk-noargs-and-seam-caller-search.out sections 4 and 5 are defective and retained. The script cd'd into <worktree>/lang for section 3 and never returned, so sections 4 and 5 ran with CWD = lang/ and their pathspecs (lang/tests/integration/) resolved to lang/lang/… and matched nothing. The "NONE" printed by section 4 was an artifact of the CWD, not a result — this is precisely the cycle-01 defect repeating. Sections 1–3 of 0039 ran from the repo root and are sound. The corrected re-run is 0040-REDO-seam-caller-and-wrap-assertion-search.out, which prints its own CWD as a guard. Do not cite 0039 §4–5.
3. Artifact index numbers are not contiguous (0017 jumps to 0036). capture.sh derives the next index from grep -c . MANIFEST.tsv, and several of my commands were multi-line shell scripts passed to bash -c, so the command field written to the manifest contained embedded newlines and inflated the line count. No data is affected — verify-manifest.sh reports 0 altered, 0 missing — but the numbering gap is an instrument artifact, not a lost artifact. Subsequent captures put multi-line scripts in files (probe07.sh, thunk07b.sh, …, stored alongside this file) and pass a single-line command.
4. The parent-commit context measurement (0043) is stamped with the wrong SHA. The parent tree was materialised with git archive bc2f26d^ | tar -x -C /tmp/el07parent rather than a worktree (worktree creation is owned by the parent agent for this run), and /tmp/el07parent is not a git repo, so capture.sh was invoked from /tmp/rerun-v1-07 and stamped bc2f26d. The subject of that measurement is bc2f26d^. The script defends against a bad extraction by hashing codegen.el and test_compiler.el against the bc2f26d^ blobs before measuring (both MATCH, printed in the artifact). Read 0043's provenance as "taken during the bc2f26d run", not "measured at bc2f26d". The same applies to the parent arm of the timing artifacts 0036/0037.

Inputs stored alongside the evidence

probe07.sh, time07.sh, time07b.sh, thunk07.sh, thunk07b.sh, parent07.sh are the scripts whose captured output is the evidence. They are inputs, not artifacts, and are not in MANIFEST.tsv. probe07.sh in particular is a reconstruction — see divergence detail A.