feat(genesis/A11): Reconciliation Loop architectural pattern + selection rule

[danielmeppiel]

Closes #6.

TL;DR

Adds A11 RECONCILIATION LOOP to the Tier-3 catalogue (sibling of A8 ALIGNMENT LOOP, not a sub-entry) and B11 FOLD-BY-DEFAULT to the Tier-2 catalogue as its queue-level policy primitive. Wires both into the Step-3 selection rule so the architect persona surfaces the pattern when the intent describes a queue of items each driven to a terminal state under non-determinism.

Recognition, not invention. The agentic instance of the Kubernetes Operator pattern (CoreOS 2016) and the SRE control loop (Beyer et al. 2016); ancestral lineage in PDCA (Shewhart 1939) and the cybernetic loop (Wiener 1948); form-analogue REST (Fielding 2000) for the discipline-over-substrate framing.

What this PR changes

flowchart LR
  C1[architectural-patterns.md] --> C1A[+ A11 entry full template]
  C1 --> C1B[+ catalogue table row]
  C1 --> C1C[+ A8 SEE ALSO one liner]
  C1 --> C1D[+ A11 in selection heuristic]
  C1 --> C1E[+ composition diagram in How-tier3-compose section]
  C2[design-patterns.md] --> C2A[+ B11 FOLD-BY-DEFAULT entry]
  C2 --> C2B[+ Where-each-pattern-lives row]
  C2 --> C2C[+ selection-heuristic row]
  C3[SKILL.md] --> C3A[+ Step-3 trigger phrases for A11]

Loading

Three files, additive only, no existing pattern reshaped.

Why a sibling (A11), not a sub-entry of A8

A8 and A11 are both loop-shaped but their composes-list, anti-pattern set, and selection trigger phrases overlap by <30%:

• A8 = single-target convergence. One artifact, N rounds, one steward. "Iterate on this PR description until reviewers approve."
• A11 = queue-of-targets convergence. Per-item bounded loop, per-item stop-predicate from the system of record, cross-item interlocks (single-writer per item), fold-vs-defer policy at the queue level. "Drive N bug-flagged issues each to merged-PR state."

A sub-entry would force readers to skim A8 to find A11. The sibling slot is the honest signal that an architect is making a different call. A8 gets a one-line SEE ALSO; the catalogue table lists both at the same level.

Substrate-portability claim (made precise)

A11 needs only three baseline primitives from the harness:

1. Sub-agent dispatch -- the runner can spawn a fresh-context sub-agent per item.
2. Persistent state -- a write surface that survives across spawns and runner re-entry.
3. Completion signal -- each sub-agent returns a structured verdict.

Vendor sugar (Codex /goal, Claude Code /loop + /goal, Copilot CLI issues #2129 and #3364) packages the re-entry contract; not a prerequisite. We tested this by building the worked example in Copilot CLI specifically because it ships neither -- the discipline rendered without the sugar. Single worked example, framed honestly as our own workbench, not a population claim.

Composition with adjacent patterns

A11 is rarely deployed bare. It nests with three other catalogue entries along independent axes -- this is the production-deployment shape:

flowchart TB
  EV[event trigger] --> GATE
  GATE[A10 capability and sandbox gate] --> RUN
  RUN[A11 runner reads state table B4] --> SPAWN
  SPAWN[spawn per item sub-agent C2 plus C4] --> WORK
  WORK[item work]
  WORK --> A8[per item A8 ALIGNMENT LOOP if item is a draft]
  WORK --> A1[per item A1 PANEL if multi lens decision]
  A8 --> SP
  A1 --> SP
  WORK --> SP[S4 stop predicate from system of record]
  SP -->|terminal| UPD[update state table]
  SP -->|non terminal within budget| SPAWN
  SP -->|budget exhausted| HC[B10 human checkpoint]
  UPD --> POL{B11 fold or defer}
  POL -->|fold| RUN
  POL -->|defer| OUT[external queue]
  POL -->|all terminal| DONE[queue closed]
  DONE --> AUDIT[A10 audit surface]
  HC --> AUDIT
  OUT --> AUDIT

Loading

Three composition axes, each independently optional:

Axis	Pattern	Fits when	Adds
Outer wrap	A10 GOVERNED OUTER LOOP	event-triggered, audited, capability-gated	sandbox, audit surface, bounded service-account authority
Per-item drop-in	A8 ALIGNMENT LOOP or A1 PANEL	item is a draft (A8) or a multi-lens decision (A1)	per-item convergence or deliberation, independent of queue mechanics
Queue policy	B11 FOLD-BY-DEFAULT	follow-ups emerge during convergence	fold-by-default vs defer, decided at queue level

Start with A11 alone; add A10 when moving to unattended deployment; add A8 or A1 per item when the per-item work warrants the upgrade. Each addition is a separate decision.

Anti-patterns named in the entry

• LOOP WITHOUT STOP-PREDICATE -- terminates on token exhaustion, not on a deterministic gate
• DEFER-BY-DEFAULT -- the queue grows faster than it drains (B11 inverts this)
• DRIFT WITHOUT REASSERTION -- iteration re-derives state from in-context recall instead of the table
• MULTI-WRITER PER ITEM -- two sub-agents touch the same item without an interlock
• QUEUE-LEVEL INTERLOCK -- single-writer per queue is wrong grain; collapses fan-out
• A8 MISCAST AS A11 / A11 MISCAST AS A8 -- pick the right loop

Maps cleanly to handbook ch19 #10 (Not Fixing Primitives), #14 (Cost Runaway), #17 (Persona Drift).

Worked example

batch-bug-shepherd in microsoft/apm -- the first concrete realization we built while developing this pattern, not external industry validation. Ran a saga in Copilot CLI (no /goal, no /loop) driving ten bug-flagged issues to seven merged PRs across two waves. The signal isn't the conversion rate; it's that the discipline rendered, terminated, and produced merged artifacts on a substrate exposing only the three baseline primitives.

Out of scope (deliberate)

• New worked-example file under skills/genesis/examples/. Real-world reference is named in the entry; a longer walkthrough is a follow-up if wanted.
• pattern-tradeoffs.md row for A8-vs-A11. Discriminator lives in both entries' bodies; a formal matrix row is a follow-up if architects report ambiguity in practice.
• Renaming A8 or renumbering the catalogue.

Companion artifact

A handbook treatment landed in parallel: danielmeppiel/agentic-sdlc-handbook#8 introduces a Mode axis (static / style / sugar) and adds Reconciliation Loop as the first style-mode entry to ch18. Same convergent framing; different audience (handbook is reader-facing prose, this PR is architect-facing reference).