chore: archive the advisor calibration harness beside its results

[voorhs]

Archives the calibration harness beside the results it produced, and makes this experiment self-contained: it no longer references an AutoIntent branch anywhere.

Why

These seven scripts (plus their unit test) lived in deeppavlov/AutoIntent under scripts/. They were removed from that repo while preparing AutoIntent#291 for merge:

1. They are validation instruments, not library code — and AutoIntent's scripts/ is load-bearing, since check-schema.yaml runs python -m scripts.generate_json_schema_config from it.
2. test_calibration_tracker.py imported calibrate_advisor from scripts/, which is not a Python package. That import was one of the mypy failures blocking #291.

No branch references

The first draft of this PR asked AutoIntent to preserve feat/issue39-calibration-scripts forever, because reproduce.sh pinned to it. That was the wrong shape — an experiment should not hold another repo's branch lifecycle hostage — and it was also already too late:

$ git ls-remote https://github.com/deeppavlov/AutoIntent 'refs/heads/*issue39*'
$ git ls-remote https://github.com/deeppavlov/AutoIntent 'refs/pull/348/head'
b38f3c3a8612e177e76fe94552180ef43555665c        refs/pull/348/head

The branch is gone. The commit is not, and never will be: GitHub keeps every commit that belonged to a PR under refs/pull/<N>/head, merged or closed or neither. So the pin is now a commit plus the PR that owns it:

AUTOINTENT_COMMIT="${AUTOINTENT_COMMIT:-b38f3c3a8612e177e76fe94552180ef43555665c}"
AUTOINTENT_PR="${AUTOINTENT_PR:-348}"   # head of AutoIntent#348 (closed)

reproduce.sh fetches that commit by name, falls back to refs/pull/$AUTOINTENT_PR/head, falls back to a full fetch, then checks out detached. Both fetch paths were verified against the live remote from an empty repo.

reproduce.sh now runs this copy

Previously deferred to you as a judgement call; taken now, because leaving it was what made the branch load-bearing. reproduce.sh executes harness/ by absolute path and uses the AutoIntent checkout for one thing only: the advisor library and the venv around it. That is the single dependency this experiment should have, since the advisor is implemented in AutoIntent and must not be implemented anywhere else.

Making that work needed three path fixes, all of which are consequences of the files no longer living in <autointent>/scripts/:

File	Change
run_calibration_banking77.sh	Split REPO_ROOT into HARNESS_DIR (here) and AUTOINTENT_DIR (the venv). Relative *.yaml preset paths now resolve against HARNESS_DIR first — the cwd is someone else's repo now, so PRESETS=coverage_preset.yaml would otherwise silently miss.
run_phase2_isolated.sh	Same split; calls the sweep driver by absolute path and forwards AUTOINTENT_DIR.
test_calibration_tracker.py	Was broken by the verbatim copy. Its sys.path bootstrap pointed at parents[2]/"scripts", which does not exist in this repo — the test could not import its own subject. Now points at its directory.

Plus four text-only path corrections in docstrings/help (scripts/… → harness/…). The advisor API surface the harness uses is untouched.

The pin is exact, not approximate

README.md credits 85848f2 with producing results/. The pinned b38f3c3a differs from it only under scripts/, which this experiment no longer reads:

$ git diff --stat 85848f27 b38f3c3a
 scripts/calibrate_advisor.py               |  90 ++++-
 scripts/phase1b_metadata_counterfactual.py | 191 +++++++++
 scripts/phase3_reduce_to_fit.py            | 284 +++++++++++++
 scripts/render_issue39_tables.py           | 116 ++++++
 scripts/run_calibration_banking77.sh       |   4 +-
 scripts/run_phase2_isolated.sh             |  38 ++

src/autointent/ is identical between the two, so the pinned library is literally the one that produced the archived numbers.

Moving the pin forward, when you want to

The pin sits behind dev deliberately. AutoIntent#291 renames autointent._advisor to the public autointent.advisor, renames inspect → estimate and stats_from_dataset_obj → dataset_stats, moves PreflightError to autointent.advisor, and flips Pipeline.fit(preflight=…) to default "off" — so bumping AUTOINTENT_COMMIT past it needs code edits here, not just a new SHA. #291 also changes advisor behaviour, so a re-run against a newer pin is a new measurement rather than a reproduction of results/. harness/README.md spells out both halves.

Not changed

ISSUE_COMMENT.md still says "AutoIntent @ feat/feasibility-check 85848f2" — it is an archived transcript of a comment already posted to issue #39, and it already carries the commit. Left verbatim rather than rewritten after the fact.

Base

Targets feat/advisor-calibration-laptop6gb rather than main, since experiments/advisor-calibration-laptop6gb/ only exists there — this stacks on #40.