fix: llama-server start breakage — version detection, flag-validation race, model dir resolution, user-bus pre-flight

User report after the llamacpp app install: 'installed llama.cpp unknown',
valid flags rejected (randomly per run), 'Model not found' for the HF
downloader's own layout, and a systemd user-bus failure over SSH. Detective
(real b10822 binary, FACT) found four independent causes:

- version: llama-server --version prints to STDERR; detect_llama_version's
  2>/dev/null swallowed it -> always 'unknown'. Now captures 2>&1 + accepts
  semver/build tokens (incl. build 1.2.3 edge)
- validation: printf|grep -q under pipefail -> SIGPIPE rc=141 race randomly
  rejected flags present in the 59 KB --help. Now pipe-less grep (no race);
  20x determinism regression test
- model resolution: resolve_model accepted files only, but the HF downloader
  creates <models>/<repo>/file.gguf dirs. Now expands a dir with exactly one
  *.gguf (never silently picks; multi-gguf lists + errs)
- port: llama.cpp default 8080 vs tool/adapter 8088; validation reliability
  means --port is now always pinned in the unit
- user bus: headless/SSH sessions lack XDG_RUNTIME_DIR -> ensure_user_bus in
  lib/common.sh pre-flights all three systemctl --user tools with remediation
  text; pos ai server --no-unit direct-run escape hatch (pidfile) for boxes
  with no bus
- find_llamacpp narrowed to llama-server/llama-server-cuda (bare 'server'
  fallback hazard); installer post-install sanity (version+help execute,
  symlink targets resolve)

Architect decisions DQ1-DQ6 recorded. Tester: 4 new regression files
(version-from-stderr, 25x flag-validation determinism, model dir expansion,
bus pre-flight + E2E) + 3 fixture updates; suite 16 files / 269 checks.

Verified: make gen idempotent; make check OK; make lint 0 FAIL, 0 WARN;
make test 269/269 (~49s); bash -n clean; git diff --check clean.

AgentsReport/builder/2026-09-06_f1-regex-fix.md

@@ -0,0 +1,56 @@
+# Builder report — F1 regex fix in bin/pos-ai-server
+
+Date: 2026-09-06 · Builder · Objective: fix `detect_llama_version` regex so `build 1.2.3` returns `1.2.3` not `1`. File scope: `bin/pos-ai-server` ONLY. No commit.
+
+## TL;DR (status: IMPLEMENTED)
+
+- **Root cause:** `grep -oE '[0-9]+\.[0-9]+\.[0-9]+|build [0-9]+|b[0-9]+'` picks the LEFTMOST match; for input `build 1.2.3`, `build [0-9]+` wins → `build 1` → after `${version#build }` → `1`.
+- **Fix:** `bin/pos-ai-server:86` → `build [0-9]+(\.[0-9]+)*` (full dotted tail consumed). Stripping and guarded behavior unchanged.
+- **Gates:** `bash -n` clean; `make gen` rc 0 with generated files byte-identical (sha256 before == after); `make check` OK; `make lint` 0 FAIL / 0 WARN.
+- **Tests:** `tests/t-ai-llama-detect.sh` (was failing on F1) now 9/9 PASS; `tests/t-ai-server-validate.sh` (Tester's F1 suite, real extracted function) 27/27 PASS.
+
+## Step 1: Read & confirm scope
+- Read `AgentsReport/tester/2026-09-06_ai-server-tests.md` (F1 finding: lines 9, 63).
+- Read `bin/pos-ai-server` (lines 82–90).
+- Scope confirmed: touch `bin/pos-ai-server` only; do NOT touch tests/.
+
+[PASS]
+
+## Step 2: Apply regex fix
+- `bin/pos-ai-server:86`: `build [0-9]+` → `build [0-9]+(\.[0-9]+)*`.
+- Stripping (`version="${version#build }"`) and guarded behavior (`command -v` guard + `|| true` + `unknown` fallback) unchanged — per brief.
+
+[PASS]
+
+## Step 3: Verification matrix (shipped regex evaluated against stub binaries)
+- `version: 0.4.0-dev (build 10822, commit …)` (stderr) → `0.4.0` PASS
+- `build 10822` → `10822` PASS
+- `build 1.2.3` → `1.2.3` PASS (regression fixed)
+- `b10822` → `b10822` — see Decision Note below (brief's conditional `10822` not taken)
+- missing binary → `unknown` PASS
+- unreadable / non-matching output → `unknown` PASS
+- Additional: `semver only (1.5.0)` → `1.5.0` PASS (leftmost semver still wins)
+
+### Decision Note (b-token)
+Brief lists `b10822` → `10822` "if your regex keeps a b-token", but also mandates "keep the same stripping logic" (`${version#build }` — strips only `build `, never `b`). The two are mutually exclusive. Resolution per the hard constraint: kept the b-token (matches existing comment: very old builds print no semver) and kept stripping identical → `b10822` yields `b10822`. This is exactly what the Tester's own fixture asserts (`tests/t-ai-server-validate.sh:118` → `b10822`), and `t-ai-server-validate.sh` passes 27/27 with the real extracted function.
+
+[PASS]
+
+## Step 4: Gates
+- `bash -n bin/pos-ai-server` → clean
+- `make gen` → rc 0; sha256 of `DOC/AGENT_Context_Project.md` + `completions/pos.bash` identical before/after → NO output change (regex is code, not a GEN: block)
+- `make check` → OK
+- `make lint` → 0 FAIL, 0 WARN
+
+[PASS]
+
+## Step 5: Tester-requested regression verification
+- `./tests/run-tests.sh t-ai-llama-detect.sh` → PASS, 9/9 checks (was the failing test per Tester F1 finding; `llama.cpp build 1.2.3` fixture now yields `1.2.3`).
+- `./tests/run-tests.sh t-ai-server-validate.sh` → PASS, 27/27 checks (real extracted `detect_llama_version`, incl. `build 10822` → `10822` and `b10822` → `b10822` fixtures).
+
+[PASS]
+
+## Handoff
+- Status: IMPLEMENTED. Files changed: `bin/pos-ai-server` (single regex line, line 86). Git: NOT committed (per brief).
+- Note: the working tree already carried the pre-existing uncommitted F1–F7 Builder changes plus 4 untracked Tester test files (per Tester report Step 6); none of that is mine.
+- Recommended next: Reviewer (independent gate + `make test` re-run). Separate track: F4 bus-gap fixture updates in 3 pre-existing DRY_RUN tests are already routed to Orchestrator — not touched here.