feat(#121): contaminated shared-context experiment

[Agastya191]

What

Adds the contaminated shared-context experiment (#121), which injects a spurious signal into the shared workspace itself instead of through a peer's turn, isolating the shared-context channel from peer conformity. It adds:

• a board_seed param on run_committee (blackboard.py) that drops an artifact -- a pre-screen flag, a "prior similar cases were all X" note, a leaked proxy metric -- onto the board before the member loop, not attributed to any agent, kept out of the committed map, and consuming no member slot (so seed_turn indexing is untouched). A board seed attributed to a committee member now raises, since that would leak a peer.
• a context flag on schema.Turn marking that artifact as ambient workspace context, so _visible_turns keeps it visible even in isolated mode -- isolation is meant to hide peers, not the shared workspace, which is the whole point of the isolated control.
• contaminated_context.py, which runs the 2x2 (present/absent x shared/isolated) and reports the anchoring: present minus absent is the artifact's pull, and shared minus isolated (shared_amplification) is the extra pull only the shared channel adds. The twin condition is held constant across all four arms on purpose, so the only thing differing between present and absent is the artifact, not the AgentView.condition label.
• composition with the existing referee (score_shortcut reliance) and the blind-metric probe (blind_metric_uptake, with the artifact as the hidden proxy), rather than reinventing either.

Plain run_committee behavior is unchanged when board_seed isn't passed -- the new context field defaults to False and every existing path stays byte-identical, same as before.

Note: this touches the shared Turn schema (a new context field) rather than overloading the existing seeded flag, because the cascade's seed_turn is a peer utterance that must stay hidden in isolated mode while a workspace artifact must not -- flagging in case you'd rather key isolated-mode visibility off an agent_id convention instead. One thing I left alone: the two visibility arms share a run_id (shared isn't in it), same as run_cascade already does, and meta["shared"] disambiguates them -- say the word if you'd rather I make run_id unique.

Testing

tests/test_contaminated_context.py:

• the 2x2 runner produces the anchoring ordering (shared > isolated > absent) on a small set
• referee flags reliance on the artifact, and blind-metric composition gives a positive uptake delta when the artifact leaks a proxy
• a condition-keying backend produces zero anchoring, proving the condition label can't confound the contrast
• ground truth stays untouched (the absent arms recover the correct answer), the run is deterministic, and the guards fire (empty cases/committee, agent-id collision, duplicate case_id, None implied answer, implied == correct)

board_seed contract tests appended to tests/test_blackboard.py:

• the artifact lands before the first agent and is visible in isolated mode too
• it never leaks into committed, never shifts seed_turn's slot index, and neither mutates the caller's Turn nor aliases it across two runs
• a board seed or pre_hook context turn attributed to a member is rejected

Suite: 492 -> 519 passing locally, offline mock path only (couldn't run ruff check in this environment, so it still needs a local lint pass before merge).

[sebasmos]

Reviewed the design (not merging, this one is yours @Agastya191). This is a clean and valuable addition: board_seed as ambient workspace context, flagged context=True and visible even in isolated mode, is exactly the right way to separate the shared-context channel from peer conformity. It also makes a first-class harness feature of what our clean-A result already saw empirically, a wrong "system pre-screen" flag is adopted by the weaker tier ~0.33 of the time (control 0.00) while a bare peer is resisted (see #141).

Two notes:

1. This and feat: reasoned rationale for cascade seed turns (addresses #115) #149 both extend run_committee's signature, so whichever lands second will need a trivial rebase, worth coordinating with @MohShahin.
2. Please confirm the context=True turn is excluded from the per-agent committed map and from the conformity / cascade-onset counters, so it can never be counted as a member turn.

Nice work.

[Agastya191]

confirmed context = True is excluded so it never can be counted as a member turn. Coordinated With Mohammad as well. Merging to main now.

[sebasmos]

Thanks @Agastya191, that resolves my note (the context=True turn excluded from the committed map and the conformity / cascade-onset counters). Good to have coordinated with @MohShahin on the run_committee signature. Go ahead and merge.