Eve library

Whole-tree discovery, not whole-tree teaching

Guide

Revision 1

On this page

September 7, 2026 UTC. Source pin: 1ceae52c7896a59ac74a4429d410f3218f0b91c6. Implementation base: 79814f43c6a7271747dd652ebbaa15ac54c2659a.

The repository index preserves all 140,123 tracked Git entries in 43 root groups. NUL-delimited discovery retains dot-directories, unusual filenames, equal-content files in distinct contexts, executable modes, one symlink and one gitlink. Nothing follows symlinks, fetches external submodules/LFS payloads, executes project code or reads local runtime data. Git blob sizes describe the committed bytes, not external asset sizes.

The collector projects 3,324 package and 3,505 project JSON manifests, all 58 registry declarations and 88 code-root declarations. It verifies original manifest bytes against their Git object identities. It retains all 44 non-root package manifests outside the declared pnpm patterns, all ten repeated names, all 2,088 unclassified paths and the registry/navigation name differences. These are review inputs, not automatic defects, deployment evidence or a resolved workspace graph. Raw scripts and dependency URLs are not copied into the index.

The 130,861 paths outside the documentation inventory make its boundary explicit. This is a file count, not a new count of meaningful units. Source-unit teaching remains 26 reviewed, 10,018 structural and 103,300 unassigned documentation units; non-Markdown semantic extraction, teaching assignments, effective dependency resolution and the final Domains audit remain unfinished. No slide, guide, source pin, speech plan or product behavior is changed by this discovery work.

Evidence#

  • Fourteen new tests pass for exact tree parsing, symlink/gitlink handling, malformed metadata, path hints, restricted glob semantics and exclusions, safe manifest projection, contextual name matches, missing/changed/extra shards, byte-verified Git reads and mixed/missing source rejection. Controlled fixtures are not live workspace-resolution evidence.
  • Sixteen existing inventory tests pass unchanged. Teaching-report freshness passes while reporting the incomplete teaching state. Ruff passes for presentation tooling.
  • A fresh full recollection matches every generated discovery file. The measured check took 8.57 seconds and peaked at 480,316 KiB RSS. It is metadata/source-accounting work, not an Nx/build/browser/model gate.
  • The read-only verifier checks exact inputs and receipt bytes, every root-shard hash and all manifest projections, then independently compares all discovered paths with Git's --name-only output. The two paths must agree exactly; counts alone cannot conceal an omitted root or file.

The existing CPU narrator continued separately during this bounded discovery work, with more than 10 GiB available in its resource samples. Its changing MP3s and manifest are not inputs or results of this receipt. Full narration/browser verification and publication remain their own in-progress task. No optical, normal-speed listening or human-reader review is claimed here.

Reproduce#

sh
python tools/presentations/inventory-repository.py --check
python -m unittest discover -s tools/presentations/tests -p 'test_repository_inventory.py'
python -m unittest discover -s tools/presentations/tests -p 'test_inventory.py'
python tools/presentations/check-center-coverage.py --check
node docs/presentations/presentation-center/verification/repository-discovery-2026-09-07/verify-evidence.mjs

Use the prepared dependencies and current resource checks. The source pin is unchanged; newer worktree or remote changes require a deliberate future source review, not an automatic replacement of this evidence boundary.