feat(v1.99 PR-17): package-install target detection + preflight deepening

.github/workflows/ci-update-canonization.yml

@@ -228,22 +228,47 @@ jobs:
           echo "G3-U2 PASS — current=$CURRENT target=$TARGET"
 
       # ------------------------------------------------------------------
-      # G3-U4 — preflight enforcement: all five checks must appear in the
-      # plan's Preflight block. We don't require every check to pass —
-      # CI environments don't have nftband running — but every check
-      # MUST be evaluated and reported.
+      # G3-U4 — preflight enforcement: all seven checks must appear in the
+      # plan's Preflight block (PR-16 P-1..P-5 + PR-17 P-6..P-7). We don't
+      # require every check to pass — CI environments don't have nftband
+      # running — but every check MUST be evaluated and reported.
       # ------------------------------------------------------------------
-      - name: G3-U4 — preflight reports all 5 checks
+      - name: G3-U4 — preflight reports all 7 checks
         shell: bash
         run: |
           set -Eeuo pipefail
-          for name in authority_nftban service_nftband_active artifact_version_file dependency_nft state_no_stale_in_progress; do
+          for name in \
+              authority_nftban \
+              service_nftband_active \
+              artifact_version_file \
+              dependency_nft \
+              state_no_stale_in_progress \
+              rebuild_recovery_available \
+              install_origin_coherent; do
             grep -q "$name" /tmp/dryrun.out || {
               echo "::error::G3-U4 FAIL: preflight check '$name' missing from plan"
               exit 1
             }
           done
-          echo "G3-U4 PASS — all 5 preflight checks reported"
+          echo "G3-U4 PASS — all 7 preflight checks reported"
+
+      # ------------------------------------------------------------------
+      # G3-U4-deepen (PR-17) — recovery planning metadata must be rendered
+      # so apply (PR-18) has explicit input about rollback reachability.
+      # ------------------------------------------------------------------
+      - name: G3-U4-deepen — recovery plan metadata rendered
+        shell: bash
+        run: |
+          set -Eeuo pipefail
+          grep -q "Recovery plan" /tmp/dryrun.out || {
+            echo "::error::G3-U4-deepen FAIL: plan has no Recovery block"
+            exit 1
+          }
+          grep -q "Mechanism" /tmp/dryrun.out || {
+            echo "::error::G3-U4-deepen FAIL: Recovery block missing Mechanism"
+            exit 1
+          }
+          echo "G3-U4-deepen PASS — recovery planning metadata present"
 
   summary:
     name: Update Canonization summary

cmd/nftban-installer/update_dryrun.go

@@ -69,25 +69,30 @@ func runUpdateDryRun(ctx context.Context, exec executor.Executor, sf *state.Stat
 	}
 	log.PhaseEnd("Detect")
 
-	// 2. Update-specific preflight (P-1 through P-5).
+	// 2. Install origin — declared flag wins; fall back to package-manager
+	// probe (PR-17) so package installs that don't re-pass --rpm/--deb
+	// still get a correct plan.
+	origin := detectInstallOrigin(cfg)
+	if origin == "" {
+		origin = update.DetectInstallOrigin(exec, log)
+	}
+
+	// 3. Update-specific preflight (P-1..P-7 — PR-16 + PR-17).
 	log.Phase("Preflight")
-	pre := update.Preflight(exec, log)
+	pre := update.Preflight(exec, log, origin)
 	log.PhaseEnd("Preflight")
 
-	// 3. Version detection. sourceDir is empty in the package-install case;
-	// the update package treats that as non-fatal for PR-16 (package-install
-	// target detection lands in PR-17).
-	current, target, err := update.DetectVersions(exec, cfg.sourceDir, log)
+	// 4. Version detection — source tree wins; package manager is the
+	// fallback for package-install hosts (PR-17).
+	current, target, err := update.DetectVersions(exec, cfg.sourceDir, origin, log)
 	if err != nil {
 		log.Error("update dry-run: version detection failed: %v", err)
 		return state.ExitFailed
 	}
 
-	// 4. Install origin — drives which apply path PR-18 will use.
-	origin := detectInstallOrigin(cfg)
-
-	// 5. Assemble + render plan.
+	// 5. Assemble plan + attach PR-17 recovery metadata (planning-only).
 	plan := update.BuildPlan(pre, current, target, origin)
+	plan.AttachRecovery(update.BuildRecoveryPlan(exec))
 	plan.Render(os.Stdout)
 
 	// 6. Write a copy of the plan JSON to the state dir for audit/history

internal/installer/update/plan.go

@@ -77,39 +77,78 @@ type Plan struct {
 	// this abstract ("reuse rebuild pipeline"); PR-18 fills in concrete
 	// steps.
 	Actions []string `json:"actions,omitempty"`
+
+	// Recovery carries planning-only metadata about the rollback route that
+	// PR-18's apply step will rely on. PR-17 populates this but does NOT
+	// execute any recovery — that's PR-18 scope per INV-U-002.
+	Recovery *RecoveryPlan `json:"recovery,omitempty"`
+}
+
+// RecoveryPlan is PR-17 metadata describing the rollback path apply will
+// follow on failure. It is DESCRIPTIVE only; no state mutation in PR-17.
+type RecoveryPlan struct {
+	// Available reports whether a rollback to the prior known-good state
+	// is possible without operator intervention.
+	Available bool `json:"available"`
+
+	// Mechanism is the rollback mechanism that would be used — currently
+	// "rebuild" (validate-in-namespace + flush+load + v1.96 recovery), the
+	// only canonized recovery surface today.
+	Mechanism string `json:"mechanism"`
+
+	// Notes carries human-oriented context that helps an operator reason
+	// about recovery without digging into code (e.g. "prior state file
+	// present and terminal", "no prior install_state — fresh recovery").
+	Notes []string `json:"notes,omitempty"`
 }
 
 // PlanSchemaVersion is the current wire contract for Plan. Consumers should
 // reject plans with unexpected schema versions.
 const PlanSchemaVersion = "1.99.0"
 
 // DetectVersions reads the current installed version from the FHS VERSION
-// file and the target version from the source-tree VERSION (for source
-// installs) or from the package metadata path.
+// file and the target version from:
 //
-// For PR-16 we support source-install detection; package-install target
-// detection is handled in PR-17.
-func DetectVersions(exec executor.Executor, sourceDir string, log *logging.Logger) (current, target string, err error) {
+//   - <sourceDir>/VERSION when sourceDir is non-empty (source installs)
+//   - the package manager (rpm -q / dpkg -s) when origin is "rpm"/"deb"
+//     and sourceDir is empty (PR-17 scope)
+//
+// Missing target is non-fatal — the plan carries a warning downstream.
+// Missing current IS fatal — we must know where we are before planning
+// any transition (INV-U-002 rollbackability).
+func DetectVersions(exec executor.Executor, sourceDir, origin string, log *logging.Logger) (current, target string, err error) {
 	current = readCurrentVersion(exec, log)
 
-	// Source install: target VERSION lives at <sourceDir>/VERSION.
+	// Priority 1: source tree VERSION (takes precedence over package query
+	// so a caller can test with a source tree even on a package-install host).
 	if sourceDir != "" {
 		tPath := filepath.Join(sourceDir, "VERSION")
 		if data, rErr := exec.ReadFile(tPath); rErr == nil {
 			target = strings.TrimSpace(string(data))
+			log.Debug("update: target version from source tree %s = %s", tPath, target)
 		} else {
 			log.Debug("update: source VERSION not readable at %s: %v", tPath, rErr)
 		}
 	}
 
+	// Priority 2 (PR-17): package manager query. Only consulted when source
+	// detection did not yield a result. We pass in `origin` rather than
+	// auto-detecting inside this function so the caller stays authoritative
+	// about the install-origin decision.
+	if target == "" && (origin == "rpm" || origin == "deb") {
+		if t, qErr := DetectPackageTarget(exec, origin, log); qErr != nil {
+			log.Warn("update: package target query failed: %v", qErr)
+		} else if t != "" {
+			target = t
+			log.Debug("update: target version from package manager (%s) = %s", origin, target)
+		}
+	}
+
 	if current == "" {
 		return current, target, fmt.Errorf("update: cannot detect current version (no VERSION file)")
 	}
 	if target == "" {
-		// PR-16: target detection for package installs lands in PR-17.
-		// For now, missing target is non-fatal — the plan records it and
-		// the preflight flags it as a warning.
-		log.Info("update: target version not yet detected (package-install detection lands in PR-17)")
+		log.Info("update: target version not detected (no source tree + no package ownership)")
 	}
 	return current, target, nil
 }
@@ -172,6 +211,20 @@ func (p *Plan) Render(w io.Writer) {
 		fmt.Fprintln(w, "")
 	}
 
+	if p.Recovery != nil {
+		fmt.Fprintln(w, "  Recovery plan (metadata only — apply lands in PR-18):")
+		availMark := "available"
+		if !p.Recovery.Available {
+			availMark = "NOT available — operator intervention may be required"
+		}
+		fmt.Fprintf(w, "    Rollback    : %s\n", availMark)
+		fmt.Fprintf(w, "    Mechanism   : %s\n", displayString(p.Recovery.Mechanism))
+		for _, n := range p.Recovery.Notes {
+			fmt.Fprintf(w, "      · %s\n", n)
+		}
+		fmt.Fprintln(w, "")
+	}
+
 	if len(p.Actions) > 0 {
 		fmt.Fprintln(w, "  Actions (dry-run — no mutation):")
 		for _, a := range p.Actions {
@@ -209,11 +262,22 @@ func BuildPlan(pre *PreflightResult, current, target, origin string) *Plan {
 	}
 	if target == "" {
 		p.Warnings = append(p.Warnings,
-			"target version not detected (package-install target detection lands in PR-17)")
+			"target version not detected (no source tree + no package ownership)")
+	}
+	if origin == "" {
+		p.Warnings = append(p.Warnings,
+			"install origin not detected — pass --source, --rpm, or --deb to disambiguate")
 	}
 	return p
 }
 
+// AttachRecovery decorates the plan with the PR-17 recovery metadata.
+// Separate from BuildPlan so callers without an exec handle (tests) can
+// still build a plan without synthesizing recovery state.
+func (p *Plan) AttachRecovery(r *RecoveryPlan) {
+	p.Recovery = r
+}
+
 // displayVersion returns a placeholder for empty versions so the rendered
 // table doesn't print blank cells that look like a bug.
 func displayVersion(v string) string {

internal/installer/update/preflight.go

@@ -61,7 +61,10 @@ type PreflightResult struct {
 // Called by phaseDetect when cfg.mode == "upgrade".
 //
 // The function is READ-ONLY. It must never mutate host state.
-func Preflight(exec executor.Executor, log *logging.Logger) *PreflightResult {
+//
+// origin is the operator-declared install origin ("rpm", "deb", "source",
+// or ""). PR-17 uses it for the new P-7 coherence check.
+func Preflight(exec executor.Executor, log *logging.Logger, origin string) *PreflightResult {
 	res := &PreflightResult{Passed: true}
 
 	// P-1 authority — the existing authority.Detect would give us a full
@@ -150,9 +153,116 @@ func Preflight(exec executor.Executor, log *logging.Logger) *PreflightResult {
 		logCheck(log, c)
 	}
 
+	// P-6 (PR-17) rebuild recovery available — the planning-only check
+	// that tells apply (PR-18) whether a clean rollback is reachable.
+	// We do NOT execute recovery here; we just report whether the
+	// prerequisites exist: terminal prior state + ip nftban table + nft
+	// binary. If any is missing, apply must treat rollback as
+	// operator-assisted rather than automatic (INV-U-002).
+	{
+		ok := rebuildRecoveryAvailable(exec)
+		c := PreflightCheck{
+			Name:     "rebuild_recovery_available",
+			Passed:   ok,
+			Severity: "warning",
+		}
+		if !ok {
+			c.Detail = "rebuild recovery may require operator intervention on failure — review before apply"
+		}
+		res.Checks = append(res.Checks, c)
+		logCheck(log, c)
+	}
+
+	// P-7 (PR-17) install origin coherent — the declared --rpm / --deb /
+	// --source flag must match the host's actual install origin. A
+	// mismatch (e.g. --rpm on a DEB host) is a hard warning: apply would
+	// route through the wrong delivery model. PR-17 detects this early.
+	{
+		declared := origin
+		detected := DetectInstallOrigin(exec, log)
+		coherent := declared == "" || detected == "" || declared == detected
+		c := PreflightCheck{
+			Name:     "install_origin_coherent",
+			Passed:   coherent,
+			Severity: "warning",
+		}
+		if !coherent {
+			c.Detail = "declared origin (" + declared + ") does not match detected origin (" + detected + ")"
+		}
+		res.Checks = append(res.Checks, c)
+		logCheck(log, c)
+	}
+
 	return res
 }
 
+// rebuildRecoveryAvailable reports whether the prerequisites for automatic
+// rollback via rebuild recovery are present:
+//
+//   - the prior install_state is terminal (COMMITTED / DEGRADED / FAILED_*)
+//   - ip nftban table exists (authority held)
+//   - nft binary is present (handled by P-4; we still check here so this
+//     predicate is self-contained for callers outside the full preflight)
+//
+// PR-17 scope: this is metadata only. PR-18 will wire it into the apply
+// path via INV-U-002.
+func rebuildRecoveryAvailable(exec executor.Executor) bool {
+	if !exec.NftTableExists("ip", "nftban") {
+		return false
+	}
+	if r := exec.Run("sh", "-c", "command -v nft >/dev/null 2>&1"); r.ExitCode != 0 {
+		return false
+	}
+	// Prior state must be readable OR absent (absent = fresh recovery is
+	// fine). If present but non-terminal (in-progress), recovery is not
+	// clean.
+	const p = "/var/lib/nftban/state/install_state"
+	if !exec.FileExists(p) {
+		return true
+	}
+	data, err := exec.ReadFile(p)
+	if err != nil {
+		return false
+	}
+	s := trim(string(data))
+	if s == "" {
+		return true
+	}
+	// Same terminal-state list as isStaleInProgress but inverted.
+	switch s {
+	case "COMMITTED", "DEGRADED",
+		"FAILED_SSH_UNKNOWN", "FAILED_AUTHORITY_ABORT", "FAILED_RENDER",
+		"FAILED_REBUILD", "FAILED_NO_FIREWALL", "FAILED_TAKEOVER":
+		return true
+	}
+	return false
+}
+
+// BuildRecoveryPlan derives the PR-17 recovery metadata from exec state.
+// Planning-only: no mutation, no recovery execution.
+func BuildRecoveryPlan(exec executor.Executor) *RecoveryPlan {
+	r := &RecoveryPlan{
+		Mechanism: "rebuild",
+	}
+	r.Available = rebuildRecoveryAvailable(exec)
+	const p = "/var/lib/nftban/state/install_state"
+	if !exec.FileExists(p) {
+		r.Notes = append(r.Notes,
+			"no prior install_state — a fresh recovery is clean (no rollback target)")
+	} else {
+		data, err := exec.ReadFile(p)
+		if err == nil {
+			r.Notes = append(r.Notes,
+				"prior install_state = "+trim(string(data)))
+		}
+	}
+	if !r.Available {
+		r.Notes = append(r.Notes,
+			"preconditions for automatic rollback unmet — apply will require operator assist on failure")
+	}
+	return r
+}
+
 // isStaleInProgress returns true if install_state exists AND is not a
 // terminal state (COMMITTED / DEGRADED / FAILED_*). We treat an in-progress
 // marker as a caution.