Date: 2026-07-29 Claim: Hermetic cross-worktree build produces identical content to default monorepo build Old rung: in_tree New rung: live_local
Build worktree: /home/travis/ProjectAmp2/.amp/worktrees/docs (branch amp/docs, commit 3f1f647)
Source root: /home/travis/ProjectAmp2 (branch main, commit 69397b8)
The docs repo is a separate git repository; the monorepo is a workspace where subdirectory repos are co-located but gitignored
The original resolve-root.mjs used DOCTRINE.md as the stack marker and git status --porcelain (including untracked files) for the clean-checkout check. Two problems:
`DOCTRINE.md` is untracked in the monorepo — no clean checkout can ever contain it
`git status --porcelain` shows untracked files (??), but the workspace root legitimately contains untracked sub-repo directories that do not affect build content
Fix (commit a79dffd):
Stack marker changed to STACK_COMPLETION.md (committed AND already a required build dependency)
Porcelain check changed to git status --porcelain -uno (tracked files only)
All commands below were executed on 2026-07-29 from the docs worktree at commit 3f1f647.
$ npm run build # exit 0
build provenance: 3f1f64782311e05835153e3456862a4aa8a58114
229 real docs mirrored from 23 projectsOutput saved to /tmp/atlas-default.json and /tmp/index-default.html.
$ cd /home/travis/ProjectAmp2 && git stash push -m "temp: stash .gitignore for hermetic build" -- .gitignore
Saved working directory and index state On main: temp: stash .gitignore for hermetic build
$ git -C /home/travis/ProjectAmp2 status --porcelain -uno
(empty — clean)
$ ATLAS_SOURCE_ROOT=/home/travis/ProjectAmp2 npm run build # exit 0
build provenance: 69397b8b73489e0f5a54a43cb42dfe3e2efdbe8e
229 real docs mirrored from 23 projects
$ cd /home/travis/ProjectAmp2 && git stash pop
(restored)Output saved to /tmp/atlas-explicit.json and /tmp/index-explicit.html.
atlas.json comparison (Python json.load, strip commit + generatedAt, deep equality):
Default build commit: 3f1f64782311e05835153e3456862a4aa8a58114
Explicit build commit: 69397b8b73489e0f5a54a43cb42dfe3e2efdbe8e
Default generatedAt: 2026-07-29
Explicit generatedAt: 2026-07-28
Default page count: 229
Explicit page count: 229
RESULT: atlas model content is IDENTICAL (excluding provenance fields)
index.html comparison (regex-strip 40-char hex hashes + generatedAt, string equality):
RESULT: index.html content is IDENTICAL (excluding provenance)
Dark-factory phases (read from both atlas.json outputs):
perceive: live_local ← STACK_COMPLETION.md
decide: GAP
act: live_local ← STACK_COMPLETION.md
measure: live_local ← STACK_COMPLETION.md
RESULT: IDENTICAL in both buildsOnly commit and generatedAt differ — these are provenance fields that correctly reflect which source tree each build was invoked against. All content-addressed page CIDs, dark-factory phase derivations, cross-doc links, and outline anchors are byte-identical.
$ ATLAS_SOURCE_ROOT=/nonexistent npm run build # exit 1: "ATLAS_SOURCE_ROOT does not exist: /nonexistent"
$ ATLAS_SOURCE_ROOT=/tmp npm run build # exit 1: "ATLAS_SOURCE_ROOT is not a stack root (STACK_COMPLETION.md not found): /tmp"All 7 regression cases pass (node test-source-root.mjs, exit 0).
$ node test-dark-factory.mjs
✓ dark-factory: 4 phases, order [perceive → decide → act → measure], decide=GAP, no rung literals in home.mjs
exit 0
$ node test-source-root.mjs
✓ source-root: 7 cases (absent, invalid, non-git, dirty, valid, missing-source, no-artifact-leak) all pass
exit 0
$ npm run build
exit 0: dist/index.html, dist/atlas.json, dist/bend/A second independent run from the same worktree at a later commit, confirming identity still holds.
Build worktree: /home/travis/ProjectAmp2/.amp/worktrees/docs (branch amp/docs, commit 8d6f53a)
Source root: /home/travis/ProjectAmp2 (branch main, commit 69397b8)
$ npm run build # exit 0
build provenance: 8d6f53acd78359ab5188074854bbc43cdd1e0216
229 real docs mirrored from 23 projects$ cd /home/travis/ProjectAmp2 && git stash push -m "temp: stash .gitignore for hermetic build" -- .gitignore
Saved working directory and index state On main: temp: stash .gitignore for hermetic build
$ git -C /home/travis/ProjectAmp2 status --porcelain -uno
(empty — clean)
$ ATLAS_SOURCE_ROOT=/home/travis/ProjectAmp2 npm run build # exit 0
build provenance: 69397b8b73489e0f5a54a43cb42dfe3e2efdbe8e
229 real docs mirrored from 23 projects
$ cd /home/travis/ProjectAmp2 && git stash pop # exit 0, restoredatlas.json comparison (JSON.parse, strip commit + generatedAt, string equality):
Baseline commit: 8d6f53acd78359ab5188074854bbc43cdd1e0216
Explicit commit: 69397b8b73489e0f5a54a43cb42dfe3e2efdbe8e
Baseline date: 2026-07-29
Explicit date: 2026-07-28
Baseline pages: 229
Explicit pages: 229
All page CIDs identical: true
RESULT: atlas model content is IDENTICAL (excluding provenance fields)
index.html comparison (regex-strip 40-char hex hashes + generatedAt, string equality):
Baseline length: 6752713
Explicit length: 6752713
RESULT: index.html content is IDENTICAL (excluding provenance)
Dark-factory phases (extracted from embedded model JSON in both index.html outputs):
perceive: live_local ← STACK_COMPLETION.md
decide: GAP
act: live_local ← STACK_COMPLETION.md
measure: live_local ← STACK_COMPLETION.md
RESULT: IDENTICAL in both builds (JSON.stringify equality)Only commit and generatedAt differ — these are provenance fields that correctly reflect which source tree each build was invoked against.
$ node test-dark-factory.mjs # exit 0
✓ dark-factory: 4 phases, order [perceive → decide → act → measure], decide=GAP, no rung literals in home.mjs
$ node test-source-root.mjs # exit 0
✓ source-root: 7 cases (absent, invalid, non-git, dirty, valid, missing-source, no-artifact-leak) all passThird independent run, completing the interrupted verification from a surviving worktree.
Build worktree: /home/travis/ProjectAmp2/.amp/worktrees/docs (branch amp/docs, commit 277acac)
Source root: /home/travis/ProjectAmp2 (branch main, commit 69397b8)
$ npm run build # exit 0
build provenance: 277acacd27e27a0b9172283f7c746c7e0bc3582f
229 real docs mirrored from 23 projects$ cd /home/travis/ProjectAmp2 && git stash push -m "temp: stash .gitignore for hermetic cross-worktree build" -- .gitignore
Saved working directory and index state On main: temp: stash .gitignore for hermetic cross-worktree build
$ git -C /home/travis/ProjectAmp2 status --porcelain -uno
(empty — clean)
$ ATLAS_SOURCE_ROOT=/home/travis/ProjectAmp2 npm run build # exit 0
build provenance: 69397b8b73489e0f5a54a43cb42dfe3e2efdbe8e
229 real docs mirrored from 23 projects
$ cd /home/travis/ProjectAmp2 && git stash pop # exit 0, restoredatlas.json comparison (JSON.parse, strip commit + generatedAt, string equality):
Baseline commit: e7d7b502feb961978e2c011e2489fb84e4d633e6
Explicit commit: 69397b8b73489e0f5a54a43cb42dfe3e2efdbe8e
Baseline date: 2026-07-29
Explicit date: 2026-07-28
Baseline pages: 229
Explicit pages: 229
All page routes/links/backlinks identical: true
RESULT: atlas model content is IDENTICAL (excluding provenance fields)
index.html comparison (regex-strip 40-char hex hashes + dates, string equality):
Baseline length: 6752713
Explicit length: 6752713
RESULT: index.html content is IDENTICAL (excluding provenance)
Dark-factory phases (extracted from embedded model JSON in both index.html outputs):
perceive: live_local ← STACK_COMPLETION.md
decide: GAP
act: live_local ← STACK_COMPLETION.md
measure: live_local ← STACK_COMPLETION.md
RESULT: IDENTICAL in both builds (JSON.stringify equality)$ git diff --exit-code HEAD # exit 0 (clean before build)
$ npm run build # exit 0
$ node test-dark-factory.mjs # exit 0
$ node test-source-root.mjs # exit 0All four checks pass against the committed revision.
Fourth run, after the provenance churn fix (7617d0a: derive build provenance from source commit, not HEAD). This is the first run where the combined check — including post-build git diff --exit-code HEAD — passes without requiring a follow-up dist/ rebuild commit.
Build worktree: /home/travis/ProjectAmp2/.amp/worktrees/docs (branch amp/docs, commit e2afa93)
Source root: /home/travis/ProjectAmp2 (branch main, commit 69397b8)
Provenance fix: Default build now uses git log -1 --format=%H -- . ":(exclude)dist" to derive provenance from the last source-file commit (7617d0a), not HEAD. This breaks the self-referential churn cycle where committing dist/ changes HEAD, which changes the embedded hash.
$ git diff --exit-code HEAD # exit 0 (clean)
$ git rev-parse HEAD
e2afa9395aa6e9c56bdf7fd583e87888b3bfbd2b$ npm run build # exit 0
build provenance: 7617d0a2265cab2669765bc621634859e1229568
229 real docs mirrored from 23 projectsOutput saved to /tmp/atlas-baseline-e2afa93.json and /tmp/index-baseline-e2afa93.html.
$ cd /home/travis/ProjectAmp2 && git stash push -m "temp: stash .gitignore for hermetic cross-worktree build" -- .gitignore
Saved working directory and index state On main: temp: stash .gitignore for hermetic cross-worktree build
$ git -C /home/travis/ProjectAmp2 status --porcelain -uno
(empty — clean)
$ ATLAS_SOURCE_ROOT=/home/travis/ProjectAmp2 npm run build # exit 0
build provenance: 69397b8b73489e0f5a54a43cb42dfe3e2efdbe8e
229 real docs mirrored from 23 projects
$ cd /home/travis/ProjectAmp2 && git stash pop # exit 0, restoredOutput saved to /tmp/atlas-explicit-e2afa93.json and /tmp/index-explicit-e2afa93.html.
atlas.json comparison (JSON.parse, strip commit + generatedAt, string equality):
Baseline commit: 7617d0a2265cab2669765bc621634859e1229568
Explicit commit: 69397b8b73489e0f5a54a43cb42dfe3e2efdbe8e
Baseline date: 2026-07-29
Explicit date: 2026-07-28
Baseline pages: 229
Explicit pages: 229
Content identical (excl. provenance): true
RESULT: atlas model content is IDENTICAL (excluding provenance fields)
index.html comparison (regex-strip 40-char hex hashes + dates, string equality):
Baseline length: 6752713
Explicit length: 6752713
RESULT: index.html content is IDENTICAL (excluding provenance)
Dark-factory phases (extracted from both atlas.json outputs):
perceive: live_local ← STACK_COMPLETION.md
decide: GAP
act: live_local ← STACK_COMPLETION.md
measure: live_local ← STACK_COMPLETION.md
RESULT: IDENTICAL in both builds (JSON.stringify equality)$ git diff --exit-code HEAD # exit 0 (pre-build clean)
$ npm run build # exit 0
$ node test-dark-factory.mjs # exit 0: 4 phases, order correct, decide=GAP
$ node test-source-root.mjs # exit 0: 7 cases all pass
$ git diff --exit-code HEAD # exit 0 (post-build clean — provenance churn resolved)All five checks pass. This is the first run where the post-build cleanliness check passes without intervention — the provenance churn fix (deriving from source commit, not HEAD) eliminates the self-referential dist/ modification cycle.
The monorepo workspace is not a single committed tree — sub-repos are co-located checkouts, not committed content. This means:
The ATLAS_SOURCE_ROOT validation checks tracked-file cleanliness (-uno), not full workspace purity
A truly hermetic build against a clean monorepo checkout is structurally impossible because the monorepo gitignores all sub-repos
The stash-and-restore of .gitignore was needed because the monorepo had a tracked-file modification unrelated to build content
This is an honest gap. The cross-worktree identity holds against the developer machine's workspace, which is the real substrate. A CI-grade hermetic build would require either a monorepo that commits all content or a manifest of sub-repo commits to clone.