Feat/context overlay sidecar

[FScholPer]

Why

An agent working in one S-CORE repository knows nothing about what was decided
in another session, in another module, last week. The knowledge exists — it is
just not selectable. Storing context is not the hard part; selecting the
context that matters for the task at hand is.

This PR adds that selection layer to context-discipline. It is local,
deterministic, and Git-native: no service, no embedding model, no network call.

What an agent gets

get_prior_context(task_text, current_nodes) returns the reasoning records
from other sessions that are relevant to the current task, plus the score
factors of the ones it rejected, so an empty result is diagnosable instead of
silent.

Relevance is one number per candidate:

score = (0.6 * semantic + 0.4 * structural + bonus) * recency * live_ratio

• semantic — token overlap between the current task and the prior record.
  Unicode-aware, function words filtered; without that filter, German prose
  matched on die/und/wird and produced a measured false positive.
• structural — the share of the prior record's grounded nodes that lies
  inside the code-graph neighbourhood of the current task. This is what lets
  "same pattern, different module" match at all.
• live_ratio — how much of a record's ground still exists in the graph;
  floored so that build/config decisions are not permanently unreachable.
• recency — off by default; retention already drops records after 90
  days, and on a small corpus age does not discriminate.

Selection is by rank, not by an absolute threshold: candidates above the
measured noise_floor and within rank_gap_ratio of the best one. Every
lookup writes an attention record with all factors, so any decision the layer
made can be replayed afterwards.

Why the numbers are what they are

They are measured, not chosen. Each pair of runs was one task that should
match and one that must not:

• noise_floor = 0.037 is the geometric mean of the worst relevant score and
  the best control score after tokenization was fixed.
• rank_gap_ratio = 0.35 is the value that still keeps the weakest relevant
  prior once structural raises the top of the scale.

In the most recent real run, all five relevant priors were returned and both
control records were rejected, with a factor of 36 between them.

Scope and boundaries

• Everything stays in .score-local/ on the developer's machine. Nothing is
  shared between developers in this PR.
• Prior reasoning is rendered inside an explicitly untrusted block: it is data
  recorded by other sessions, never instructions to follow.
• record_outcome now requires verdict to be exactly pass or fail, with
  the justification in the new rationale field. This is a breaking change
  for callers that passed prose verdicts.
• structural only contributes when current_nodes contains module or
  directory paths; the tool schema now says so at call time, and the response
  reports focus_size so an agent can see that its input carried no structural
  signal.

What is deliberately not in here

Exploration sampling, IDF weighting (it makes every score depend on corpus
composition, so a recorded decision stops being reproducible), and a
specificity cap on focus expansion (measured, it lost more signal than it
saved). Each was tried against the real graph and rejected on evidence.

Frank Scholter Peres frank.scholter_peres@mercedes-benz.com, Mercedes-Benz Tech Innovation GmbH
Provider Information

AI Assisted by Devin