Feedback wanted: can a map an agent writes about its own repo be trusted?

[tuna781]

dspec rests on one bet, and it is the part I want argued with rather than praised.

The map in .ds/ is prose an agent writes about the repository it is working in — what each
feature is, where it lives, the rules it must not break, and why it is the way it is. Nothing
verifies it. There is no parser, no schema, no fingerprint, no drift check. 0.2.0 deleted all of
that on purpose. What keeps the map honest is three instructions the agent is given and nothing
else:

• the code wins — when .ds/ and the code disagree, the map is what is wrong, and fixing it
  is part of the change that broke it;
• nothing is described unread — a session updating the map may not rewrite a description of
  code it has not opened;
• what could not be settled is written down as unsettled, so the confident half can be
  believed.

That is the whole mechanism. It is in templates/bootstrap.md and templates/memory.md, about
240 lines of prose between them, and the program that installs them is ~900 lines of dependency-free
TypeScript.

The three questions I actually want answered:

1. Is "the code wins, fix the map" enough? Or does a knowledge layer need verification — a
   fingerprint, a check in CI — before a team can rely on it? I removed verification because the
   version that had it spent its budget proving the map was stale instead of making it useful, but
   that is one person's read of one tool.

2. Does ## Unsettled survive contact with a real team? The idea is that a map which marks
   the edge of what it knows can be trusted about the rest. The failure mode I expect is that it
   degrades into a backlog nobody reads, or that agents write nothing there because admitting
   doubt reads as failure.

3. Where does this break at scale? It is built for a repository with tens of features. At a
   few hundred, the index stops being scannable and the "read the index first" instruction starts
   costing more than the search it replaces. I don't know where that line is.

What I am not asking for: a line-by-line review, or a verdict on whether the idea is good.

If you want to try it in five minutes rather than read about it, demo/shop is a 15-file
storefront with a map committed. Ask your agent why does checkout reject my second discount code?
in there, then ask the same thing with .ds/ deleted, and compare which files each session opens.

Disagreement is more useful to me than agreement here — if you think the premise is wrong, that is
the comment I want.