Audit the rest of the codebase; fix a live traversal hole

[jadoon200]

Covers what the M0-M3 and control-layer audits did not: the queueing core, the adaptive
controllers, the Pareto machinery, the price client, and the deployed container. Four
defects, one a security hole in the live image.

1. Path traversal in the deployed dashboard (security)

The SPA catch-all joined the request path onto dist and served whatever it found.
Percent-encoded traversal survives URL normalisation and reaches the handler intact:

/..%2f..%2frequirements-serve.txt   -> served the file
/%2e%2e/%2e%2e/%2e%2e/etc/hostname  -> served the file

Verified against the real deploy image under uvicorn, which also ran as root — so this
was an arbitrary read of the whole container. No credentials exist to steal (the project is
keyless by design), but source, manifests and system files were readable.

Render's edge rejected the encoded forms with a 400, so the live site was not exploitable.
That is an accident of the CDN, not a control, and the README's docker run had no cover.

Fixed: candidates are resolved and confined to dist (closing symlink escapes and
absolute-path injection too), and the image drops to uid 10001. The deploy-image CI lane now
asserts both against the running container.

2. The budget pacer was steering on a frozen capacity trace

adaptive.py kept its own copy of the actuation-delay rule. Fixing the clock-restart bug
in apply_actuation_delay left that copy behind. On a plan that moves every step — what
C5/C6 emit — the mirror reported a flat 13 replicas across 40 steps while the simulator
provisioned 13 → 24. 36 of 40 steps disagreed, so the PI loop corrected against
violations that never happened.

Fixed by extracting ActuationTracker, with the batch form built on it so they cannot
diverge. Tests agree them over 200 random plans, and one fails if the rule is re-implemented
outside simulator.py — this being the second drift of that same logic.

No conclusion moves. Only C5/C6 rows shift, in both directions and by small amounts. Q4
is unchanged at 9 better / 3 tie / 6 worse of 18 with the below-19:1 pattern intact; the GPU
frontier verdict is unchanged at 0 of 8.

3. Hypervolume measured the wrong staircase

It swept from cost zero and credited each strip with the dearer point's violation rate,
when only the cheaper point is affordable there. One point at cost 5 with 0.2 violations
against a reference of 10 returned 8.0 where the rectangle is 4.0 — and the error can
reverse a ranking, not just inflate both sides. The old test compared two curves where it
cancelled. Nothing published used the function.

4. The two fidelities were documented as the same system

utilisation_target derates the utilisation model but not the queue, so at 0.8 the queue
runs at rho = 0.8 where the utilisation model reports a step at its limit. Harmless — every
published result uses the utilisation model, and Erlang-C validation sets the target to 1.0 —
but the docstring asserted something false about the code.

Verified sound

ACI matches the published BACC update including its sign under miscoverage; pareto_front is
a correct two-objective sweep; cost_at_matched_violation reports unreachable targets rather
than substituting the nearest point; the price parser skips non-positive prices rather than
defaulting them. OData filters are now escaped — not a live injection path, but an apostrophe
would have returned the wrong SKU's price.

Gate

ruff clean; mypy strict clean; 201 tests; make validate-simulator GREEN.