feat(G4): the session remembers what it was told

[pmanko]

Delivery gate G4 — layered context, stacked on G3b (#66). Companion Hub prompt: pmanko/med-agent-hub#21.

What a session now carries

An instruction that mattered on turn two mattered on turn five, and the only thing accumulating in a session was the list of instructions — text the prompt explicitly ranks below the current one, so a correction had to be repeated to survive. Three bounded layers now ride on every revision request:

Layer	Rule
Guidance	≤20 active entries, delivered verbatim in pin order
Verified examples	≤3 kept versions from this session, same source and catalog, ranked by word overlap then newest turn then stable id
Relevant failure	the one prior failure, and only while it is still the last thing that happened

Decisions worth reviewing

• Verbatim, always. The gateway does not tidy what a person wrote — the wording is the instruction. Mutation-tested: adding a .strip() turns a test red.
• Append-only. Unpinning or replacing appends a lifecycle event and leaves the text; a turn's evidence keeps meaning what it meant when it ran. Past the cap the oldest stop being delivered and nothing is forgotten.
• Precedence is stated, not implied. Position reads as authority to a model, so guidance says it outranks the retained history but not the current instruction, and that a later entry wins a conflict. Examples and failures are labelled evidence, not commands.
• Empty layers are absent, not empty objects a model pays tokens to read past.
• Negotiated. Catalyst deploys ahead of the Hub, so a Hub that has not advertised catalyst.query.session-context.v1 is not sent the layer. Handing it over to be silently ignored is worse than withholding it — the guidance would look delivered and change nothing.
• Counted before sending. Overflow is a refusal, not a truncation that would drop the pinned guidance while the turn still looked like it honoured it. A profile naming no exact tokenizer cannot be counted; a character-count substitute under-counts exactly this dense JSON.

Evidence

• 328 passed, 3 skipped (gateway) · 692 passed (hub, in the companion PR).
• 30 new tests. Every guard mutation-tested: normalising the text, dropping omissions, exceeding three examples, feeding a turn its own answer, ignoring overlap ranking, delivering empty layers, uncapped delivery, deleting instead of appending, sharing guidance across sessions, excluding the output reserve, and tolerating a missing tokenizer — each turns a test red.
• Kept on purpose: the current turn is excluded from its own examples at the call site even though it cannot appear in prior turns yet; that mutant survives at the service level and the rule is covered where it is decided.

sessionContext is optional in the revision contract and writerOutcome optional on the turn, so readers written before Phase 1 keep working.