Pin EigenScript v0.35.1, drop the record_history opt-out, re-point planted fault 1

[InauguralPhysicist]

Closes #24

EigenScript v0.35.1 carries #829, the fix for #827 (unbounded temporal
assignment history, armed on dead code — the bug that froze a box while
building this view). This bumps the pin, removes the workaround it made
unnecessary, and — the actual point — re-points the planted fault that the
fix silently retired.

Preflight (before the pin moved)

Built a worktree of the v0.35.1 tag (e2bb217) and ran the whole suite
against that binary with the tree unchanged. --version is not evidence, so
the fix was confirmed in the binary directly: a live prev of loop at 400k
iterations peaks 52,992 kB on v0.35.0 and 2,944 kB on v0.35.1; the
dead-code arming repro, 15,488 kB → 3,072 kB.

Every test passed except test_bif_mem.sh, which failed exactly where
the issue predicted:

--- planted fault 1: the temporal-history opt-out removed (the real bug) ---
FAIL: the leaking build passed the memory gate — the gate can't discriminate

That is the whole issue in one line: the fault could no longer be planted.

What changed

• Pin: .devcontainer/Dockerfile EIGS_REF v0.35.0 → v0.35.1. Re-verified
  it is the only pin site in the repo (eigs.json carries no runtime dep).
• Workaround removed: _run's record_history of 0 / record_history of hist_was pair and the comment block above it.
• Planted fault 1 re-pointed (not deleted).

Removing the opt-out costs nothing measurable

Peak RSS in the bifurcation view, at 30 / 100 / 300 frames on v0.35.1:

	30	100	300
static, opt-out in	134,272 KB	134,528 KB	134,400 KB
static, opt-out out	134,656 KB	134,272 KB	134,400 KB
toggle, opt-out in	134,272 KB	134,400 KB	135,040 KB
toggle, opt-out out	134,528 KB	134,528 KB	134,912 KB

Every difference is inside the ±400 KB run-to-run noise. No regression.

Why fault 1 had to be re-pointed, with the measurement

Same tree, the old fault planted (session history explicitly ON), static leg:

runtime	30 frames	100 frames	300 frames
v0.35.0 (old pin)	304,384 KB	674,304 KB	—
v0.35.1 (new pin)	134,528 KB	134,912 KB	134,784 KB

~5.3 MB/frame before, dead flat after. The fault is not weaker on the new
runtime; it does not exist. Deleting it would have left the gate validated by
one fewer fault than it was designed around.

The replacement fault, and why this one

The new fault makes the bifurcation tick keep the frame's plotted points —
it builds the point set every frame and appends it to app state. Chosen over
the two candidates the issue floated, both of which were evaluated and
rejected on measurement rather than taste:

• the chart's chart_trim / series cap — this view never calls chart_trim
  and sets no series cap: the sweep is static data handed over once, so there
  is no cap to remove. hover_enabled (the one real per-frame guard in the
  chart config) is scanned on mouse motion, and the gate's driver sends none,
  so flipping it moves nothing.
• the sim.cap history bound — real, but it retains one two-element list per
  advanced frame: ~15 KB over the whole 300-frame toggle leg against a 25 MB
  limit. It is already covered by a planted fault in test_orbit_hist.sh.

What the chosen fault buys: it breaks the invariant this view's headline
comment actually claims ("BIFURCATION MODE DOES NOTHING PER FRAME"), it is the
same class as both retention bugs this repo has had (the unbounded sim.pts
trail, and F-DYN-13's pinned per-point allocations), and it leaves every
structural counter correct — so only the shape of the memory curve can catch
it, which is precisely the check that needed a witness.

It goes red on both RSS checks (gate run with the quiet flag lifted so the
failing checks print):

--- planted fault 1: the bifurcation tick keeps its frame's points ---
  static 30 frames: peak 176384 KB  rc=0  built=1 series=1 points=4000 markers=4
  static 100 frames: peak 279104 KB  rc=0  built=1 series=1 points=4000 markers=4
FAIL: static/100 peaked at 279104 KB, ceiling 204800 KB
  static growth over the frame sweep: 102720 KB (limit 25600 KB)
FAIL: static RSS grew 102720 KB with frame count — that is a leak, not a working set
PASS: planted fault 'framepts' is caught (per-frame retention, every structural counter still correct)

Fault 2 still bites, unchanged:

--- planted fault 2: the build-once sweep guard removed ---
  toggle 30 frames: peak 187192 KB  rc=0  built=1 series=16 points=4000 markers=64
FAIL: toggle/30 built 16 chart series — the build-once guard let go
FAIL: toggle/30 has 64 markers, want 4
PASS: planted fault 'guard' is caught (every view switch rebuilds the sweep)

Gate contract satisfied: PASS on a clean build, FAIL on every remaining fault.

  static growth over the frame sweep: 768 KB (limit 25600 KB)
  toggle growth over the frame sweep: 512 KB (limit 25600 KB)
PASS: peak RSS under 204800 KB and flat in frame count, in both the steady-state and the toggle-every-frame shape
PASS: memory gate green, both planted faults caught

Validation

Full suite green on v0.35.1 — test_lint (17 files clean, no new upstream lint
rule fires on this repo's code), test_smoke, test_lab, test_orbit_hist,
test_bifurcation, test_orbit_oracle, test_orbit_mouse, test_bif_mem.
Every gfx run was capped (ulimit -v 1500000) and headless (xvfb-run -a).

Docs updated: F-DYN-13 marked fixed upstream with the before/after numbers, and
the README's memory-gate paragraph records why a planted fault was re-pointed.

Closes #24

🤖 Generated with Claude Code