probbit persona: the individuality layer (CLI, MCP, Python, wasm)

[BitmapAsset]

probbit persona gives an agent an individual temperament that lives outside the model. A persona file (YAML subset or JSON:
traits with priors, moods with inertia, couplings, per-turn inputs, history features, habits as hard rules) + an individual's
state + a turn's inputs compile to ONE probbit-ir program, answered in process by probbit run's own instruction at fixed work.
The stance (a level per trait with exact odds, the habits in force and the ones that changed it, a refusal when the engine cannot
vouch, a why, a stance line of at most 40 estimated tokens) and the next state are canonical JSON. Format: docs/persona.md.

What is in it:

• CLI probbit persona init | turn | replay | explain | diff | lint | check | compile | describe (probbit-cli/src/persona.rs),
  a strict YAML-subset reader (yaml.rs), a strict validator (one {"error": {"code": "persona", ...}} object, exit 2), one
  canonical number rule for every digest (docs/persona.md §5.2), SHA-256 in process. No dependency added.
• MCP probbit_persona_init, probbit_persona_turn in probbit mcp (stateless; the CLI's documents).
• Python probbit.persona_init, persona_turn, persona_replay.
• Browser probbit-wasm ops 4 and 5; the playground's "Meet three individuals from one persona".
• Docs docs/persona.md, a README section "The individuality layer", rows in docs/agents.md, CHANGELOG.
• Examples examples/persona/: three fictional personas (YAML + JSON, one digest each) and their goldens.

Parity: the goldens equal an independent reference implementation's documents byte for byte (state0, all 20 workday stances,
the final state, lint), except that each engine.program digest is of a program whose IR key is probbit_ir (the reference ran
the 0.4.0 engine, key pbit_ir; the program texts are otherwise byte-identical on every turn). A further 5,400 random-input turns
(3 personas x 3 seeds x as written / values: semantic / forced onto the sampler; held, yielded, refused and fallback turns)
are equal too.

Tests (all new, probbit-cli/tests/persona.rs and probbit-wasm/tests/persona.rs): stdout_matches_the_persona_goldens (both
file forms, init / compile / 20 x turn / replay / lint), the YAML subset's accepted and refused cases, 1,000-turn replays
byte-identical across processes and against turn --state, habit properties over 2,160 turns of 12 random personas (every habit
in force re-checked, 481 bound habits broken by the habit-free twin, positional == semantic), lint (clean examples, planted
contradictions statically, at load and at run time), the evaluate bridge at judge weight 0 == the run path (625 trait answers),
refusals (no unvouched trait in the line), 600 malformed documents (never a crash, one error object), Python 3.9
(python/test_persona.py), MCP answers == CLI answers (python/test_mcp.py), wasm documents == CLI documents. The CI wasm job's
check.mjs also replays the three personas in the real browser module against the goldens.

Changed in existing files: probbit mcp lists seven tools (the two persona tools after the five; python/test_mcp.py's two
tool-list assertions updated); sys.rs gains loadavg(); usage and help list probbit persona. No other command's behaviour or
output changes; the 0.5.0 golden digests are untouched.

Checks (final sha 20c8529ad613):

• CI run 37088504110: green on ubuntu, macOS and Windows (test: cli 65, persona 10, evaluate 8, stress 11 + 1 ignored,
  acceptance 19, ir 47, core 4, wasm no_threads 3 + persona 1), the wasm job (check.mjs: each persona's 20 workday stances in the
  real module equal the goldens) and the four release targets.
• Local (Apple M4, macOS 26.5.2): cargo test --release --workspace 159 existing + 1 ignored + 11 new; the two wall-clock tests of
  probbit-decide/tests/acceptance.rs miss their 10 ms local limit only at 1-minute load 33-46 (they pass on CI).
• Clippy: no warning in any added file; the edited files' warning set is unchanged.
• Python 3.9.6: test_persona.py 5, test_mcp.py 8, test_probbit.py unchanged and passing.
• A turn in process: median 0.20-0.32 ms, p95 0.22-0.67 ms (N = 1,000 per persona, load 3.3-3.5). Binary 1,400,880 -> 1,862,432
  bytes (+461,552); wasm module 839,523 -> 1,169,601 bytes (+330,078).
• No dependency added (Cargo.lock unchanged).