v1.1: the patterns release

v1.1: the patterns release

What's new

Three guardrail patterns, extracted from the production system this repo documents, now
ship as runnable, stdlib-only code with their real tests:

• budget_governor: stops an agent from dying mid-task
  when a shared usage cap runs dry.
• proactive_compactor: keeps a long session from
  stalling on a blocking flush at rotation, and includes the three-call experiment that
  proved the tempting "background summarize" design was wrong:
  resume does not fork.
• escalation_valve: hands a hard call up to a stronger
  model with a receipts ledger instead of a cap.

Alongside the patterns, docs/RELIABILITY-WEEK.md is a new writeup
covering the hardening pass that produced them: what got proposed, what got measured
before being trusted, and what got caught only by running the thing, not by review or
tests alone. The README now leads with the patterns instead of the case
study, and each pattern links back to the reliability writeup for the receipt behind it.

Why

This repo started as a case study of a production system. The July arc was different:
instead of shipping a new feature, the goal was making what already shipped harder to
break and cheaper to run forever, and then pulling the load-bearing pieces of that
hardening pass out into something a reader could actually run, not just read about. The
patterns in this release are behavior-preserving extractions of that real logic, not
rewrites for the demo.

The numbers

python3 -m pytest patterns/ -q
................................                                 [100%]
32 passed in 0.03s

32 tests, hermetic, zero live model calls, zero writes outside a tmp directory, under 2
seconds on a clean clone. The receipts behind each pattern (the phantom-spend catch, the
per-compaction cost, the fork experiment) are in each pattern's own README and in
docs/RELIABILITY-WEEK.md.

Honest limits

These patterns are the mechanism, not the whole system. The governor only sees spend you
record into it; the compactor briefs thin long arcs and durable facts still belong in
your own memory system; the valve's consults are only as good as the question you inline.
Wiring three independent patterns together is still the caller's job, documented in
patterns/README.md, not solved by an import between
pattern directories.