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/tester/2026-09-06_busgap-fixtures.md

@@ -0,0 +1,49 @@
+# Busgap Fixture Update — 2026-09-06
+
+## TL;DR
+
+- **F4 (`ensure_user_bus`)** in `bin/pos-ai-server` now runs in `cmd_start` BEFORE the DRY_RUN return (verified in `lib/common.sh:148` + `bin/pos-ai-server:654-671`), so every unit-path `start` without a reachable user systemd bus aborts with rc 1 and no `ExecStart:` output.
+- Three pre-existing DRY_RUN test files never provided a `systemctl` stub / bus env, so 33 checks failed (26 + 3 + 4).
+- **Fix:** added a succeeding `systemctl` stub (return 0) to each affected sandbox, matching the established pattern in `t-systemd-unit.sh` and `t-ai-server-bus.sh`. No assertions weakened; bus-missing behavior remains covered in `t-ai-server-bus.sh`.
+- **Result:** full suite green — **16 files pass / 0 fail, 269 checks pass / 0 fail, 0 skip. Runtime 72s.**
+
+**PASS/FAIL totals: 269 pass / 0 fail. Defects by severity: none.**
+
+---
+
+## Step 1: Establish failure & mechanism
+
+Read `bin/pos-ai-server`, `lib/common.sh`, `tests/run-tests.sh`, the 3 affected tests, and the passing control tests (`t-systemd-unit.sh`, `t-ai-server-bus.sh`).
+
+**Verified call order:** `cmd_start` (pos-ai-server:654-655) calls `ensure_user_bus` when `NO_UNIT != 1`, i.e. **before** the `DRY_RUN` early-return at line 661. `ensure_user_bus` (common.sh:148-154) runs `systemctl --user show-environment &>/dev/null` and `err`s on nonzero. All three DRY_RUN tests drive the unit path (no `--no-unit`), so without a `systemctl` stub they abort with rc 1 and no `ExecStart:` output.
+
+## Step 2: Apply fixtures (systemctl stub returning 0)
+
+Established pattern: `printf '#!/usr/bin/env bash\nexit 0\n' > "$stubs/systemctl"` + add to `chmod +x` (exact copy of `t-systemd-unit.sh:41,46`). Also tested that this satisfies the `systemctl --user show-environment` probe without needing `XDG_RUNTIME_DIR` — the stub returns 0, so the pre-flight passes. Assertion lines untouched.
+
+- `t-ai-server-flags.sh`: added stub after nvidia-smi, extended chmod.
+- `t-unsupported-flags.sh`: added stub after first llama-server/chmod, extended chmod.
+- `t-config-precedence.sh` (Part B): added stub after nvidia-smi, extended chmod.
+
+[PASS]
+
+## Step 3: bash -n + full suite
+
+- `bash -n` on all three edited files: OK.
+- `make test`: 16 files / 269 checks, all pass.
+
+Counts per modified file: `t-ai-server-flags.sh` 28, `t-config-precedence.sh` 43, `t-unsupported-flags.sh` 19. (Prior failing counts: 26 / 4 / 3. The higher post-fix counts reflect the checks now actually running to completion instead of aborting on the bus error; the pre-existing assertion set was preserved.)
+
+[PASS]
+
+---
+
+## Handoff
+
+**Status:** TESTS_READY
+
+- **Files changed (test fixtures only):** `tests/t-ai-server-flags.sh`, `tests/t-unsupported-flags.sh`, `tests/t-config-precedence.sh`
+- **Assertions:** unchanged (only added a systemctl stub + chmod); bus-missing behavior stays in `t-ai-server-bus.sh`.
+- **Verification:** `bash -n` all three; `make test` → 269 pass / 0 fail, 72s.
+- **No production code changed; nothing committed.**
+- **Residual failures:** none. (Noted: F1-regex fix and these fixtures land in parallel; this run was green, so no rerun needed.)