docs(core): publish the left-of-line artifact — captures, side-by-side, static snapshot

[lex00]

#1084 produces two numbers. This produces the thing a reader actually looks at. Filing it separately because every issue in this group currently produces an input and none produces the output.

Deliverable

A docs page carrying a static, side-by-side rendering of what each tool knows before it runs your code, plus the committed captures and the procedure that regenerates them.

Rendering decisions already settled

Timeline mode, not Aggregated. It is the only view where the x-axis is time, so left literally means before. spicypath's own design rules reserve flow semantics for chart for exactly this reason. In Aggregated, horizontal position means nothing and the claim would be a lie about the axis.

Regex search-dim on the project's own module path. CDK's timeline lights up almost entirely and starts near t=0. chant's fold timeline stays full, because the TypeScript parser is doing real work, and nothing highlights. A busy screen with an empty highlight reads as a claim. An empty screen reads as a broken file, which is why an "empty flame graph" framing was rejected during analysis.

Static export, not a live link. spicypath exports SVG snapshots. The page must be readable without running a viewer.

Depends on

• test(core): left-of-line proof harness — what is known before the first user frame #1084 for the measurement and the capture procedure.
• spicypath FG-064 (FG-064 · Node/JS package attribution in packageOf (node_modules segments) spicypath#30). Without JS package attribution, aws-cdk-lib frames scatter across file basenames and the coloring carries no information. The comparison is unreadable until this lands.
• spicypath FG-065 (FG-065 · Timeline event markers (labeled vertical annotations on the time axis) spicypath#31) for the boundary marker. Search-dim shows which frames are yours but never marks when the boundary is crossed.

The expensive unassigned input

Both metrics need the same infrastructure expressed in chant and in CDK. The boundary claim survives a mismatch because it is structural, but trusted-computing-base bytes do not — profiling loomster against an unrelated AWS sample makes that number meaningless. Building equivalent estates is real work and it is currently assigned to nothing. Either scope it here or drop the second metric.

AC

• Committed profile captures for both tools, regenerable from a documented procedure.
• The CDK capture profiles the app subprocess rather than the CLI, since profiling the CLI measures the wrong process.
• The same infrastructure is expressed in both tools, or the second metric is dropped and that is stated.
• A static exported snapshot on a docs page, readable without running spicypath.
• The page states the structural claim and the current loomster fold ratio adjacent to each other, so capability and delivery are not blurred.
• The publish threshold is recorded before the measurement is taken, not chosen after seeing it.

Non-goals

Diff mode. Two structurally unrelated profiles do not diff meaningfully, and signed-weight comparison over disjoint call trees produces noise.

Any timing claim. If the page implies one tool is faster, the artifact has drifted from what it set out to show.

Relates #1019, #1084.

[lex00]

Everything this issue needs now exists. The harness, matched pair, committed captures, and exhibits are in PR #1120 (closes #1084). The page itself is PR #1124, deliberately draft: per the decision recorded on #1107, merging it is the publish click and the bar is real-src fold coverage, which loomster hasn't met. The page already states capability and delivery adjacently, so it stays honest whenever it ships.

Measured result on the matched pair (CRUD items API, parity-checked): project code executed — chant no (doubly pinned: no sampled project frame + the run fails unless every file reports [fold:fold]), cdk yes; definition-library TCB — chant 0 MB, cdk ~1.8 MB. The exhibits render exactly as this issue's body predicted: CDK's timeline lights up with a first-project-frame marker near t=0; chant's timeline is busy with an empty highlight and no frame to mark, and the title says 0 project fns matched so the absence reads as checked rather than broken.

[lex00]

Publish-bar re-measurement (2026-08-26), taken per the workstream rule that the orchestrator re-measures before the publish click.

Setup: isolated clone of loomster main @ 1c68202 (feat(loom-agents): staged AgentCore policy governance, #172), all @intentius/* deps at published latest (chant 0.49.0). Command: npx chant build src --lexicon aws --fold --verbose -o out.json, counting [fold:fold]/[fold:run] markers, real-src = paths excluding examples/.

Result (three-scope): real-src 37 fold / 21 run; byo examples 17/0; ops not a denominator per the recorded decision.

Versus the 0.24.0 capstone (33/17, holdouts exhaustively function-definition modules): loomster's new policy-governance code added run-side holdouts that are NOT in the should-never-fold category:

• loom-agents/policy/temporal.ts — dogwood.defaultEventSchema(...) function call as a value (the fold: call-form composite/intrinsic calls are the top remaining corpus gate — Checkout(...) ×8+forgejo, helm include/printf ×3 (41 files, sole blocker in 13 entries) #1174 blocker class), with an unresolved-identifier cascade through policy-engine.ts and agents.ts
• loom-agents/policy/generated/runtime.ts — re-export from @intentius/chant/runtime rejected as "chant's own module is not project source"
• loom-agents/policy/generated/index.d.ts — exported destructured or uninitialized declaration; a .d.ts appearing in the fold denominator at all may itself be a measurement artifact worth ruling on
• plus reverse-taint on policy/*.ts and params.ts files that would fold in isolation

The 7 composites/*.ts and lib/* holdouts remain the recognized should-never-fold set.

Conclusion: the bar as recorded on #1107 (every real-src file folds except genuinely-should-never-fold blockers) is not met today, so PR #1124 stays draft. #1174 (in progress) targets the call-as-value class directly; the runtime re-export and .d.ts-in-denominator findings look like two small engine questions worth their own issues if they reproduce after #1174.

[lex00]

Publish-bar re-measurement (2026-09-07)

Taken because the blocker class named in the 2026-08-26 re-measurement, #1174 (call-form composite/intrinsic calls), has since closed. Same protocol as before: the orchestrator re-measures before the publish click.

Setup. Fresh clone of loomster main @ 1c68202 (feat(loom-agents): staged AgentCore policy governance, #172), which has not moved since the last re-measurement. All @intentius/* deps resolved to published latest, chant 0.57.0 (was 0.49.0 in August).

Command.

npx chant build src --lexicon aws --fold --verbose -o out.json

counting [fold:fold] / [fold:run] markers, real-src = paths excluding examples/.

Result (three-scope). real-src 37 fold / 21 run (58 files); byo examples 17/0; ops is not a denominator per the recorded decision.

The headline number did not move, but the reason did

37/21 is bit-identical to August. The composition underneath is not. #1174's class is genuinely gone: loom-agents/policy/temporal.ts:53's dogwood.defaultEventSchema(), the call-as-value the August comment named, now folds. The file simply has a second, independent blocker 44 lines further down, and it dominates the same cascade:

temporal.ts:97:50 - function call as a value is not foldable: String(...)

from

const spendBudgetDoc = `No session spends past ${String(SPEND_BUDGET_CENTS)} cents in ${SPEND_WINDOW}.`;

String is a global built-in. It is in none of the three closed allowlists fold/fold.ts documents (registered chant authoring helper #1082, lexicon intrinsic opted into foldsAsCall #1044, project-local foldable function #1373), so the call is rejected on the general rule. Nothing on unreleased main through 0.58.0 touches that path.

Today's 21 holdouts, sorted

Recognized should-never-fold (8): the 7 composites/*.ts on "a function used as a value", plus lib/naming.ts on reverse taint from them.

One root cause and its cascade (9): policy/temporal.ts on String(...), then policy-engine.ts, agents.ts and loom-agents/outputs.ts through unresolved-identifier chains that all terminate at it, then reverse taint onto loom-agents/params.ts, policy/policies.ts, policy/stage.ts, policy/generated/index.ts and lib/params-helpers.ts.

Denominator questions, unchanged since August and still unruled (4): policy/generated/index.d.ts ("exported destructured or uninitialized declaration"), policy/generated/runtime.ts (re-export from @intentius/chant/runtime, "chant's own module is not project source"), and the local/ pair, compose.ts on the docker lexicon's env(...) with proxy.ts reverse-tainted behind it.

Counterfactual: the distance is one line

Rewriting that template span to ${SPEND_BUDGET_CENTS} and nothing else, since template spans already coerce at fold time, and re-running the same command:

real-src 46 fold / 12 run.

One loomster line clears nine files. The residue is then exactly the should-never-fold set (7 composites plus lib/naming.ts) and the four denominator questions above. No new blocker appears behind it.

Verdict

The bar recorded on #1107 is not met today, so PR #1124 stays draft. Nothing was committed.

Is it still the right bar

Yes in shape, and it is now close enough to be worth resolving rather than re-measuring. Three things would sharpen it:

1. Rule on the denominator. A .d.ts has no runtime and cannot execute, so counting it as a "run" file overstates execution rather than measuring it. policy/generated/* is codegen output, not hand-authored infrastructure. src/local is a docker-tier subtree that loomster builds with --lexicon docker, and its blocker, env(...), is the compose-interpolation ambient read the bar already exempts by name. The first two were flagged in August and are still unruled; the local/ pair is new to this reading. Together they are the difference between the bar being 13 files away and 9 files away.
2. String(...), Number(...), Boolean(...) in value position look like a fourth closed allowlist, in the same family as feat(core): chant's own authoring helpers (phase, output, Parameter) are unfoldable — caps real-app coverage #1082 and A call to a project-local imported function is not foldable, so the recommended params file defeats folding #1373 and cheaper than either. They are pure, they have no project identity, and here one of them is the sole gate on nine files. Worth its own issue if the maintainer agrees.
3. The loomster side is a one-line fix, and it is a fix loomster wants anyway, since the String() wrapper does nothing a template span does not already do.

With (1) ruled and either (2) or (3) landed, the bar is met and #1104 is a publish click rather than a measurement.

One thing to fix before the next re-measurement

loomster's package.json still depends on @intentius/chant-lexicon-temporal, which is deprecated at 0.57.0 and removed per #2116. It installs today with a warning and does not affect this measurement, but the next chant release that drops it will break this procedure. Worth a loomster issue to move ops onto the local runtime or the fountain lexicon so the publish-bar measurement stays runnable.