Skip to content

NU-002 Fixture corpus: author the tracer-bullet test documents #31

Closed

Description

@BasuruK

Part of #30.

Author and commit the five-document fixture corpus that A1, A2, A3, A5, A6, A7, A9, A10, A11 and
A12 all depend on. It currently exists only as text in a comment on #15 — nothing owns turning it
into files a pipeline can actually ingest.

Why this is its own story

Nine of the twelve assertions need these exact files, with exact constraints, before they can be
tested. Burying fixture-authoring inside whichever story happens to need it first (deterministic
core) risks a second, drifted copy showing up later when the watcher, the compiler, persistence,
merge gate, or the query agent each want to exercise the same corpus. One commit, one canonical
path, used everywhere downstream.

Build

Commit to fixtures/watched/:

  • auth-token-decision.md (A, the decision) — attribution in prose, not frontmatter
    (— Basuru, 2026-05-14). If the attribution were in frontmatter, A2's author_source="extracted"
    would test the easy path instead of the signature-regex path the parser actually needs.
  • rate-limiter.md (B, the entity doc) — "the limiter keys off the session store."
    Must never mention JWT, sessions being dropped, or the decision itself. If it did, A1 would
    stop testing traversal and start testing similarity.
  • cdn-token-caching.md (C, noise) — must be plausibly similar (tokens, auth,
    performance) rather than obviously unrelated, or retrieval has nothing to get wrong.
  • sessions-table-replication.md (D, the alias case) — introduces sessions table as a
    different name for the same referent as session store.

Exact text for all four is preserved verbatim in the comment on #15 — copy it, don't rewrite it;
the wording is tuned to the constraints above.

Also produce:

  • PDF E — a one-page PDF carrying the same decision as A (same Subject, author, date),
    for A11 only. Any tool that produces a real PDF is fine; the assertion is about what the parser
    finds after LlamaParse extraction, not about how the PDF was generated.

Definition of done

  • All four .md files exist at the paths above, byte-for-byte matching the text preserved in
    Tracer Bullet 01 — document to cited answer #15's comment.
  • PDF E exists, is a valid one-page PDF, and its text content (once extracted) carries the same
    Subject/author/date as A.
  • A human (or the next story's reviewer) can read B, C, and D and confirm none of them
    leaks the constraint they're supposed to test blind to.

Closes

None directly — this story asserts nothing. It is the prerequisite for A1, A2, A3, A5, A6, A7, A9,
A10, A11 and A12 all being testable at all.

Metadata

Metadata

Assignees

Labels

ready-for-agentFully specified, ready for an AFK agentwayfinder:taskWayfinding child ticket: implementation task

Projects

No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions