T07: agentic realization over a stdlib browser surface

Two decisions taken with the operator: stdlib HTML driver instead of
Playwright (F-0004), and a deterministic discovery runtime instead of a live
model. Both sit behind interfaces so the alternatives drop in later.

- html.py: stdlib DOM parse and query
- agentic.py: DiscoveryRuntime (agentic arm, ignores data-td by construction)
  and RecordedSelectorRuntime (control arm, uses the strongest identifier the
  page offers)
- browser.py: per-actor sessions over real HTTP, constructed per call so no
  actor inherits another's connection state
- cost/nondeterminism metrics recorded from the first run

F-0005 (CONCEPT_DRIFT): the H-001 result is a narrowing. Where test ids are
preserved, discovery 9/9 and recorded selectors 9/9 - the semantic action buys
nothing. Where they are dropped, discovery 2/3 and recorded 0/3. The concept
model presents semantic actions as generally superior; the evidence says
conditionally superior.

M21 and M22 added mid-task: the deciding side of the axis was N=1. M22 (field
names renamed) defeats the heuristic and is the first concrete evidence that a
live model would add capability, not just cost.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

Assistant: claude-code
Assistant-Model: opus
Assistant-Process: 1629012@bnt-lap001
Assistant-Session: 78d4fb13-8a1e-474b-87a3-9b9261c49a39

research/findings/F-0004-stdlib-driver-instead-of-playwright.md

@@ -0,0 +1,60 @@
+---
+id: F-0004
+type: framework-finding
+class: FRAMEWORK_LIMITATION
+status: open
+discovered: "2026-08-22"
+discovered_by: TD-WP-0002-T07
+workplan: TD-WP-0002
+task: TD-WP-0002-T07
+carried_to: TD-WP-0002-T10
+---
+
+# F-0004 — The browser surface is driven without a browser
+
+## Decision
+
+`TD-WP-0002-T07` names Playwright as the browser driver. It was not used.
+Instead, `src/testdriver/html.py` parses the lab's server-rendered HTML with
+`html.parser`, and `src/testdriver/browser.py` drives it over real HTTP.
+
+Reasons, in order of weight:
+
+1. **The lab's surface has no JavaScript.** It is server-rendered forms. A
+   browser engine would not change what H-001 or H-002 can be measured against —
+   the mutations that matter are DOM restructuring, rewording and field renaming,
+   all of which a parser faces in full.
+2. **Establishing Playwright is a real cost to the operator** — installation plus
+   a licence review, and the usual snag is the Chromium binaries the installer
+   pulls rather than the library licence itself. Blocking the vertical spike on
+   that was not worth it for a surface that gains nothing from it.
+3. **The stack is deliberately boring.** Zero dependencies remains true.
+
+`Driver` is a Protocol, so a Playwright implementation sits alongside this one
+without touching the kernel, the oracles or the evidence format.
+
+## What this costs — stated, not buried
+
+The driver cannot:
+
+- execute JavaScript, so **no single-page-application surface can be driven**;
+- take screenshots, so that evidence type listed in `INTENT.md` is unavailable;
+- read an accessibility tree, or observe visual layout at all.
+
+The third is the most consequential for the thesis. A real mechanical change
+often moves a control *visually* while leaving the DOM largely intact, and this
+driver is blind to that entire class. H-001's mutation set is therefore narrower
+than the hypothesis's wording implies: it tests **structural** durability, not
+**visual** durability.
+
+That should be said plainly whenever the H-001 result is quoted.
+
+## Carried to T10
+
+Two questions, neither answerable now:
+
+1. Does a Playwright driver actually change the H-001 result, or merely widen the
+   surface it can reach? Worth one experiment, not a rewrite.
+2. Is the screenshot evidence type in `INTENT.md` load-bearing, or was it listed
+   because screenshots are conventional in this space? If nothing has needed one
+   by T10, that is a candidate for compression.

research/findings/F-0005-semantic-actions-earn-their-keep-narrowly.md

@@ -0,0 +1,80 @@
+---
+id: F-0005
+type: framework-finding
+class: CONCEPT_DRIFT
+status: open
+discovered: "2026-08-22"
+discovered_by: TD-WP-0002-T07
+workplan: TD-WP-0002
+task: TD-WP-0002-T07
+hypotheses: [H-001]
+carried_to: TD-WP-0002-T10
+---
+
+# F-0005 — Semantic actions earn their keep more narrowly than claimed
+
+## The claim as written
+
+H-001, from `TestDriverImprovementLoop.md` § 3:
+
+> A semantic action survives implementation restructuring better than a recorded
+> UI interaction sequence.
+
+`INTENT.md` treats this as foundational — semantic actions are "the bridge
+between agentic exploration and deterministic crystallization".
+
+## The measurement
+
+Twelve mechanical mutations, two arms, same semantic action
+(`grant_access(Bob, R, READ)`), same surface, same oracles.
+
+| | Discovery (agentic) | Recorded selectors (control) |
+|---|---|---|
+| test ids **preserved** (9 mutations) | 9/9 | **9/9** |
+| test ids **dropped** (3 mutations) | 2/3 | 0/3 |
+
+The control arm is not a straw man: it uses stable `data-td` attributes, which is
+what a well-instrumented application provides and what good practice recommends.
+`test_the_control_arm_is_not_a_straw_man` asserts it keeps winning where those
+attributes survive.
+
+## What this actually says
+
+**Where an application is well instrumented and keeps its identifiers, the
+semantic action buys nothing.** Nine mutations, two arms, identical results. The
+conventional approach is not merely adequate there — it is cheaper, faster and
+deterministic.
+
+The semantic action earns its keep in exactly one circumstance: **when stable
+identifiers are absent or are not carried forward through a change.** That is a
+real and common circumstance — a rewrite rarely preserves test ids, and a large
+share of applications never had them — but it is much narrower than "survives
+implementation restructuring better", which reads as a general claim.
+
+## The second boundary: M22
+
+Discovery fails on M22, where form field names change (`subject_id` →
+`recipient`) with test ids dropped. The heuristic runtime scores candidates
+partly on field names, so renaming them removes a signal it depends on.
+
+This is an honest limit rather than a bug. It marks where a scripted runtime
+stops and where a model plausibly starts: the page still carries the label
+"Person" next to the field, which a model could read and a keyword heuristic
+cannot. **M22 is the first concrete piece of evidence that a live model would add
+capability rather than merely cost** — worth more than a general argument that it
+might.
+
+## Consequences
+
+1. **H-001 must never be quoted as a single rate.** Split by the test-id axis or
+   it is misleading. `research/hypotheses/H-001-semantic-action-stability.md` now
+   records the split.
+2. **The concept model overstates this.** `INTENT.md` and the Concept Model
+   present semantic actions as generally superior. The evidence says
+   conditionally superior. Classified `CONCEPT_DRIFT` — the concept should be
+   revised to match the evidence (§ 5 path 2), not the other way round.
+3. **Three mutations is still thin.** 2/3 and 0/3 are directionally clear and
+   statistically nothing. Any stronger statement needs more mutations on the
+   dropped-identifier side. Recorded rather than rounded up.
+4. See also **F-0004**: this driver tests structural durability only. Visual
+   relayout, the other major mechanical change class, is untested entirely.