Add a technical-memory supersession benchmark

[somtri]

Adds benchmarks/technical_memory.py. Net diff against main is that one file: 276 lines, no change to memory/, no change to the existing tests, no change to the README.

Why a separate benchmark

The drift suite measures personal-fact slot updates. Agent memory is increasingly used for standing engineering guidance instead, where the facts are build commands, paths, flags, versions and conventions. That content stresses the write gate differently. "The tests run with pytest -q" and "The tests run with pytest -n auto" are not logically contradictory sentences. They are two compatible-looking statements that disagree about an opaque token.

36 pairs, 18 positive and 18 negative, written in one pass before any policy was run against them. Three policies over one embedder and one fixture, so any gap is the policy's doing.

Result

policy           recall  precision   FP   FN
cosine-only        0.94       0.77    5    1
nli                0.94       1.00    0    1
nli+literal        0.94       1.00    0    1

The shipped NLI path transfers to technical memory at perfect precision. It never once forgot approved guidance. All five cosine-only false positives are restatements, for example "Use pytest -q to run the unit tests" against "The unit tests run with pytest -q" at cosine 0.974, and NLI scores every one of them at contradiction 0.002 or below. That is a clean domain-transferred demonstration of what the contradiction signal buys over a density gate, on content the drift suite does not cover.

The one shared false negative is an antonym pair carrying no code literal at all: "Test fixtures are generated at build time" against "Test fixtures are checked in and never generated", cosine 0.846, contradiction 0.021. Cosine-only catches it and both NLI policies miss it. That is the real remaining gap on this domain.

About the history on this branch

The first commit added an opt-in literal-diff gate to memory/sidecar.py. I had measured it on a 16-pair probe where two command swaps scored contradiction 0.013, and it recovered one of them.

The second commit removes it. On this held-out fixture the NLI model fires correctly on every command swap, contradiction 0.841 to 0.990, so the failure the gate targets does not generalise. It was a property of two specific sentence pairs, not of code-token swaps as a class. The gate is exactly inert here: same recall, same precision, same false negative as plain NLI.

I left both commits rather than force-pushing, since the correction is part of what the benchmark is for. The policy survives only inside the benchmark file, as a documented negative result so nobody re-derives it.

Honest limits

The fixture is self-authored by one person, so it measures transfer to one author's idea of technical memory rather than to a sampled population of real stores. 36 pairs is small. The three policies are evaluated as pure write-gate decisions over sentence pairs, not end to end through write(), so surprise gating and retention decay are out of scope here.

Reproduce with python -m benchmarks.technical_memory. Existing suite unchanged at 14 passed.