diff --git a/backend/cmd/tmctl/main.go b/backend/cmd/tmctl/main.go
index 527731c..09f0f9c 100644
--- a/backend/cmd/tmctl/main.go
+++ b/backend/cmd/tmctl/main.go
@@ -31,6 +31,7 @@ func run() error {
 
 	fs := flag.NewFlagSet(cmd, flag.ExitOnError)
 	cfgPath := fs.String("config", "", "path to book.yaml")
+	resnapshot := fs.Bool("resnapshot", false, "re-pin existing jobs to the current config snapshot (пере-перевод оплаченных чанков — явное согласие)")
 	if err := fs.Parse(args); err != nil {
 		return err
 	}
@@ -43,11 +44,16 @@ func run() error {
 
 	ctx, stop := signal.NotifyContext(context.Background(), os.Interrupt, syscall.SIGTERM)
 	defer stop()
-	ctx = obs.WithReqInfo(ctx, obs.ReqInfo{TraceID: obs.NewTraceID()})
+	// LogBodies решается один раз на допуске (privacy by default): содержимое
+	// LLM-обменов попадает в логи только при LOG_LLM_BODIES=1 И LOG_LEVEL=debug.
+	ctx = obs.WithReqInfo(ctx, obs.ReqInfo{
+		TraceID:   obs.NewTraceID(),
+		LogBodies: os.Getenv("LOG_LLM_BODIES") == "1",
+	})
 
 	switch cmd {
 	case "translate":
-		return translate(ctx, *cfgPath)
+		return translate(ctx, *cfgPath, *resnapshot)
 	case "report":
 		return report(*cfgPath)
 	default:
@@ -55,12 +61,13 @@ func run() error {
 	}
 }
 
-func translate(ctx context.Context, cfgPath string) error {
+func translate(ctx context.Context, cfgPath string, resnapshot bool) error {
 	r, err := pipeline.NewRunner(cfgPath, obs.NewLogger())
 	if err != nil {
 		return err
 	}
 	defer r.Close()
+	r.Resnapshot = resnapshot
 
 	res, err := r.TranslateOneChunk(ctx)
 	if err != nil {
@@ -147,7 +154,12 @@ func loadDotEnv(path string) {
 		if !ok {
 			continue
 		}
-		k, v = strings.TrimSpace(k), strings.Trim(strings.TrimSpace(v), `"'`)
+		k, v = strings.TrimSpace(k), strings.TrimSpace(v)
+		// Снимаем только ПАРНЫЕ обрамляющие кавычки: Trim по набору символов
+		// откусил бы легитимную кавычку в конце ключа (находка ревью).
+		if n := len(v); n >= 2 && (v[0] == '"' || v[0] == '\'') && v[n-1] == v[0] {
+			v = v[1 : n-1]
+		}
 		if os.Getenv(k) == "" {
 			os.Setenv(k, v)
 		}
diff --git a/backend/internal/config/models.go b/backend/internal/config/models.go
index 3072900..455f7b4 100644
--- a/backend/internal/config/models.go
+++ b/backend/internal/config/models.go
@@ -39,7 +39,11 @@ type Provider struct {
 	CacheTTL  string   `yaml:"cache_ttl"` // anthropic kind: "", "5m", "1h"
 	Model     string   `yaml:"model"`     // local kind: the backend's own tag
 	MaxTokens int      `yaml:"max_tokens"`
-	Timeouts  Timeouts `yaml:"timeouts"`
+	// Temperature (local kind): override запроса — ollama чтит request-
+	// температуру поверх Modelfile, и температура облачной роли молча
+	// расстроила бы локальную модель. 0 = наследовать запрос.
+	Temperature float64  `yaml:"temperature"`
+	Timeouts    Timeouts `yaml:"timeouts"`
 }
 
 // Timeouts is the retry profile per provider (профиль per-провайдер из
@@ -154,9 +158,18 @@ func LoadModels(path string) (*Models, error) {
 	return &m, nil
 }
 
-// Prices builds the ledger pricer from the config.
+// Prices builds the ledger pricer from the config. Besides the model entries,
+// each LOCAL provider's own backend tag is registered at $0: адаптер локальной
+// ноги отдаёт LLMResponse.Model = свой тег (huihui_ai/qwen3-…), и без этой
+// записи PriceFor свалился бы на платный default-якорь — бесплатные локальные
+// вызовы книжились бы по облачной цене, фантомно съедая потолки.
 func (m *Models) Prices() (*ledger.Pricer, error) {
-	prices := make(map[string]ledger.ModelPrice, len(m.Models))
+	prices := make(map[string]ledger.ModelPrice, len(m.Models)+len(m.Providers))
+	for _, p := range m.Providers {
+		if p.Kind == "local" && p.Model != "" {
+			prices[p.Model] = ledger.ModelPrice{}
+		}
+	}
 	for name, mod := range m.Models {
 		prices[name] = mod.Price.ToLedger()
 	}
diff --git a/backend/internal/ledger/pricing.go b/backend/internal/ledger/pricing.go
index f9bc7ca..5e15dca 100644
--- a/backend/internal/ledger/pricing.go
+++ b/backend/internal/ledger/pricing.go
@@ -75,9 +75,17 @@ func CostUSD(price ModelPrice, u llm.Usage) float64 {
 }
 
 // EstimateUSD is the pre-call reservation estimate: assume the whole prompt
-// misses the cache and the completion runs to maxTokens. Deliberately
-// pessimistic — the reservation is released down to the real cost at settle.
+// misses the cache — or, worse, is entirely WRITTEN to the cache (Anthropic
+// write ×1.25/×2 дороже input) — and the completion runs to maxTokens.
+// Deliberately pessimistic: реальная цена не должна превышать резерв, иначе
+// settle пробивает потолок, пропущенный на допуске. Известный остаточный
+// случай: additive reasoning-токены (xAI) биллятся ПОВЕРХ completion и оценкой
+// не покрыты — допустимый овершут ≤ одной резервации, как у донора.
 func EstimateUSD(price ModelPrice, promptTokens, maxTokens int) float64 {
-	return float64(promptTokens)*price.InputPerM/1_000_000 +
+	inPerM := price.InputPerM
+	if price.CacheWritePerM > inPerM {
+		inPerM = price.CacheWritePerM
+	}
+	return float64(promptTokens)*inPerM/1_000_000 +
 		float64(maxTokens)*price.OutputPerM/1_000_000
 }
diff --git a/backend/internal/llm/httpllm.go b/backend/internal/llm/httpllm.go
index 4a384e8..46a8033 100644
--- a/backend/internal/llm/httpllm.go
+++ b/backend/internal/llm/httpllm.go
@@ -4,6 +4,7 @@ import (
 	"bytes"
 	"context"
 	"encoding/json"
+	"errors"
 	"fmt"
 	"io"
 	"log/slog"
@@ -24,9 +25,16 @@ import (
 //     константы: захардкоженный 60-секундный per-attempt дедлайн донора убивал
 //     бы реальные чанки 63–278 с и трижды бесплатно их ретраил;
 //   - Retry-After на 429 читается и уважается (донор капил backoff на 8 с —
-//     молоток по rate-limit'ам массового прогона глав);
+//     молоток по rate-limit'ам массового прогона глав), но сам капится
+//     maxRetryAfterWait — иначе враждебный «Retry-After: 7200» вешает стадию
+//     на часы с удержанным резервом;
 //   - тело ответа читается через LimitReader (у донора кап был только в
 //     gemini-native адаптере).
+//
+// Задокументированное отступление от донора: one-shot self-heal параметра
+// reasoning_effort (модель, вернувшая 400, запоминалась и параметр срезался)
+// НЕ портирован — у нас reasoning задаётся per-стадия в конфиге с fail-fast
+// валидацией, и 400 от провайдера должен чиниться в конфиге, а не глотаться.
 
 // RetryProfile bounds one client's retry loop. Zero fields fall back to
 // defaults; the profile comes from models.yaml per provider (per-role
@@ -62,6 +70,59 @@ func (p RetryProfile) withDefaults() RetryProfile {
 // keeping a misbehaving endpoint from exhausting memory.
 const maxResponseBytes = 16 << 20
 
+// maxRetryAfterWait bounds an honoured Retry-After: the provider's hint wins
+// over our backoff schedule, but never for longer than this — a stage holding
+// a USD reservation must stay interruptible-by-timeout, not wedged for hours.
+const maxRetryAfterWait = 5 * time.Minute
+
+// retryLoop is the ONE retry engine shared by every transport (OpenAI-compat
+// and the native Anthropic adapter): exponential backoff with cap, Retry-After
+// override, jitter, ctx-done select. Один экземпляр политики — чтобы фикс
+// ретраев не «уезжал» от одного провайдера, забыв другой.
+func retryLoop[T any](ctx context.Context, profile RetryProfile, name string, log *slog.Logger, attempt func() (T, bool, error)) (T, error) {
+	var zero T
+	var lastErr error
+	for att := 0; att < profile.MaxAttempts; att++ {
+		if att > 0 {
+			shift := att - 1
+			if shift > 20 {
+				shift = 20 // защита сдвига от переполнения при больших max_attempts
+			}
+			backoff := profile.BackoffBase << uint(shift)
+			if backoff <= 0 || backoff > profile.BackoffCap {
+				backoff = profile.BackoffCap
+			}
+			if ra := retryAfterOf(lastErr); ra > 0 {
+				if ra > maxRetryAfterWait {
+					ra = maxRetryAfterWait
+				}
+				backoff = ra
+			}
+			backoff += time.Duration(rand.Intn(250)) * time.Millisecond
+			select {
+			case <-ctx.Done():
+				return zero, ctx.Err()
+			case <-time.After(backoff):
+			}
+		}
+		resp, retryable, err := attempt()
+		if err == nil {
+			return resp, nil
+		}
+		lastErr = err
+		if ctx.Err() != nil {
+			return zero, ctx.Err()
+		}
+		if !retryable {
+			return zero, err
+		}
+		if log != nil {
+			log.WarnContext(ctx, name+" attempt failed, will retry", "attempt", att+1, "max", profile.MaxAttempts, "err", err)
+		}
+	}
+	return zero, fmt.Errorf("%s: exhausted %d attempts: %w", name, profile.MaxAttempts, lastErr)
+}
+
 // openAIClient performs OpenAI-compatible /chat/completions calls with retry.
 type openAIClient struct {
 	name    string // provider label for logs/errors ("deepseek", "glm", "local")
@@ -99,11 +160,14 @@ type openAIMessage struct {
 }
 
 type openAIRequest struct {
-	Model       string          `json:"model"`
-	Messages    []openAIMessage `json:"messages"`
-	MaxTokens   int             `json:"max_tokens"`
-	Temperature float64         `json:"temperature,omitempty"`
-	Stream      bool            `json:"stream"`
+	Model     string          `json:"model"`
+	Messages  []openAIMessage `json:"messages"`
+	MaxTokens int             `json:"max_tokens"`
+	// Temperature is ALWAYS sent (no omitempty): дропнутый explicit T=0
+	// молча превращал детерминированную стадию (судья C2) в сэмплирование
+	// провайдерским дефолтом ~1.0.
+	Temperature float64 `json:"temperature"`
+	Stream      bool    `json:"stream"`
 	// Optional; omitempty keeps the plain body minimal.
 	ReasoningEffort string `json:"reasoning_effort,omitempty"`
 	ResponseFormat  any    `json:"response_format,omitempty"`
@@ -200,43 +264,9 @@ func (c *openAIClient) complete(ctx context.Context, reqBody openAIRequest) (*op
 	if err != nil {
 		return nil, err
 	}
-
-	var lastErr error
-	for attempt := 0; attempt < c.profile.MaxAttempts; attempt++ {
-		if attempt > 0 {
-			backoff := c.profile.BackoffBase << uint(attempt-1)
-			if backoff > c.profile.BackoffCap {
-				backoff = c.profile.BackoffCap
-			}
-			// A server-provided Retry-After overrides our schedule: the provider
-			// knows its rate-limit window better than our exponential guess.
-			if ra := retryAfterOf(lastErr); ra > 0 {
-				backoff = ra
-			}
-			backoff += time.Duration(rand.Intn(250)) * time.Millisecond
-			select {
-			case <-ctx.Done():
-				return nil, ctx.Err()
-			case <-time.After(backoff):
-			}
-		}
-
-		resp, retryable, err := c.attempt(ctx, payload)
-		if err == nil {
-			return resp, nil
-		}
-		lastErr = err
-		if ctx.Err() != nil {
-			return nil, ctx.Err()
-		}
-		if !retryable {
-			return nil, err
-		}
-		if c.log != nil {
-			c.log.WarnContext(ctx, c.name+" attempt failed, will retry", "attempt", attempt+1, "max", c.profile.MaxAttempts, "err", err)
-		}
-	}
-	return nil, fmt.Errorf("%s: exhausted %d attempts: %w", c.name, c.profile.MaxAttempts, lastErr)
+	return retryLoop(ctx, c.profile, c.name, c.log, func() (*openAIResponse, bool, error) {
+		return c.attempt(ctx, payload)
+	})
 }
 
 // attempt performs one HTTP call. Returns retryable=true for 429/5xx and
@@ -280,7 +310,11 @@ func (c *openAIClient) attempt(ctx context.Context, payload []byte) (*openAIResp
 
 	var out openAIResponse
 	if err := json.Unmarshal(data, &out); err != nil {
-		return nil, false, fmt.Errorf("%s decode: %w", c.name, err)
+		// 2xx с нечитаемым телом (обрыв/прокси/обрезка LimitReader): провайдер
+		// УЖЕ списал деньги. Retryable (повтор может прийти целым); типизировано,
+		// чтобы раннер на исчерпании попыток НЕ вернул резерв, а консервативно
+		// заселтлил оценку — недоучёт хуже переучёта для потолков.
+		return nil, true, &BilledDecodeError{Provider: c.name, Err: err}
 	}
 	// A 2xx is a billed call even when the model returns empty content
 	// (content filter, finish_reason=length with no text). Return it as a
@@ -305,13 +339,50 @@ func (e *HTTPStatusError) Error() string {
 	return fmt.Sprintf("%s http %d: %s", e.Provider, e.Status, e.Body)
 }
 
+// BilledDecodeError is a 2xx whose body could not be decoded: the provider has
+// billed the call, but usage/text are unknown. The runner treats an exhausted
+// retry chain ending in this error as BILLED (settle at the reservation
+// estimate) rather than releasing the reservation.
+type BilledDecodeError struct {
+	Provider string
+	Err      error
+}
+
+func (e *BilledDecodeError) Error() string {
+	return fmt.Sprintf("%s: 2xx body decode failed (call IS billed): %v", e.Provider, e.Err)
+}
+
+func (e *BilledDecodeError) Unwrap() error { return e.Err }
+
 func retryAfterOf(err error) time.Duration {
-	if se, ok := err.(*HTTPStatusError); ok {
+	var se *HTTPStatusError
+	if errors.As(err, &se) {
 		return se.RetryAfter
 	}
 	return 0
 }
 
+// toOpenAIMessages maps neutral messages onto the wire. CacheBoundary is
+// meaningless here: OpenAI-compatible caches (DeepSeek: prefix-match с 0-го
+// токена блоками по 64 токена) are automatic; the stable-prefix ORDER the
+// assembler produced is all that matters.
+func toOpenAIMessages(msgs []Message) []openAIMessage {
+	out := make([]openAIMessage, len(msgs))
+	for i, m := range msgs {
+		out[i] = openAIMessage{Role: m.Role, Content: m.Content}
+	}
+	return out
+}
+
+// jsonResponseFormat returns the response_format value for JSONOnly requests
+// (nil otherwise, so the field serializes away).
+func jsonResponseFormat(jsonOnly bool) any {
+	if jsonOnly {
+		return map[string]string{"type": "json_object"}
+	}
+	return nil
+}
+
 func parseRetryAfter(v string) time.Duration {
 	if v == "" {
 		return 0
diff --git a/backend/internal/llm/httpllm_test.go b/backend/internal/llm/httpllm_test.go
index 6cfccb9..967a1e8 100644
--- a/backend/internal/llm/httpllm_test.go
+++ b/backend/internal/llm/httpllm_test.go
@@ -3,6 +3,7 @@ package llm
 import (
 	"context"
 	"encoding/json"
+	"errors"
 	"net/http"
 	"net/http/httptest"
 	"sync/atomic"
@@ -163,7 +164,7 @@ func TestTerminal4xxNoRetry(t *testing.T) {
 		t.Fatal("terminal 4xx must fail")
 	}
 	var se *HTTPStatusError
-	if !asHTTPStatus(err, &se) || se.Status != 400 {
+	if !errors.As(err, &se) || se.Status != 400 {
 		t.Fatalf("want typed 400, got %v", err)
 	}
 	if calls.Load() != 1 {
@@ -210,17 +211,23 @@ func TestRetryOn5xxThenSuccess(t *testing.T) {
 	}
 }
 
-func asHTTPStatus(err error, target **HTTPStatusError) bool {
-	for err != nil {
-		if se, ok := err.(*HTTPStatusError); ok {
-			*target = se
-			return true
-		}
-		u, ok := err.(interface{ Unwrap() error })
-		if !ok {
-			return false
-		}
-		err = u.Unwrap()
+func TestBilledDecodeErrorOn2xxGarbage(t *testing.T) {
+	// 2xx с нечитаемым телом: ретраится, а на исчерпании отдаёт типизированную
+	// billed-ошибку — раннер по ней селтлит оценку вместо возврата резерва.
+	var calls atomic.Int32
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		calls.Add(1)
+		w.Write([]byte(`{"truncated`)) // обрыв тела через прокси
+	}))
+	defer srv.Close()
+
+	c := NewOpenAICompatClient(OpenAICompatConfig{Name: "test", BaseURL: srv.URL, Profile: fastProfile()}, nil)
+	_, err := c.Complete(context.Background(), LLMRequest{Model: "m", Messages: []Message{{Role: "user", Content: "u"}}, MaxTokens: 16})
+	var bde *BilledDecodeError
+	if err == nil || !errors.As(err, &bde) {
+		t.Fatalf("want BilledDecodeError, got %v", err)
+	}
+	if calls.Load() != 3 {
+		t.Fatalf("garbled 2xx must be retried to exhaustion, calls=%d", calls.Load())
 	}
-	return false
 }
diff --git a/backend/internal/llm/provider_anthropic.go b/backend/internal/llm/provider_anthropic.go
index df94f9f..650f0eb 100644
--- a/backend/internal/llm/provider_anthropic.go
+++ b/backend/internal/llm/provider_anthropic.go
@@ -7,9 +7,7 @@ import (
 	"fmt"
 	"io"
 	"log/slog"
-	"math/rand"
 	"net/http"
-	"time"
 
 	"textmachine/backend/internal/obs"
 )
@@ -93,12 +91,15 @@ type anthropicMessage struct {
 	Content []anthropicTextBlock `json:"content"`
 }
 
+// anthropicRequest deliberately carries NO sampling parameters: актуальные
+// Claude-модели (Sonnet 5, Opus 4.7+) отклоняют temperature/top_p с 400 —
+// управление стилем идёт промптом. Конфигная temperature стадии остаётся в
+// request-hash (детерминизм ключа), но на этот wire не попадает.
 type anthropicRequest struct {
-	Model       string               `json:"model"`
-	MaxTokens   int                  `json:"max_tokens"`
-	System      []anthropicTextBlock `json:"system,omitempty"`
-	Messages    []anthropicMessage   `json:"messages"`
-	Temperature float64              `json:"temperature,omitempty"`
+	Model     string               `json:"model"`
+	MaxTokens int                  `json:"max_tokens"`
+	System    []anthropicTextBlock `json:"system,omitempty"`
+	Messages  []anthropicMessage   `json:"messages"`
 }
 
 type anthropicResponse struct {
@@ -126,52 +127,26 @@ func (c *anthropicClient) Complete(ctx context.Context, req LLMRequest) (*LLMRes
 	if err != nil {
 		return nil, err
 	}
-
-	var lastErr error
-	for attempt := 0; attempt < c.profile.MaxAttempts; attempt++ {
-		if attempt > 0 {
-			backoff := c.profile.BackoffBase << uint(attempt-1)
-			if backoff > c.profile.BackoffCap {
-				backoff = c.profile.BackoffCap
-			}
-			if ra := retryAfterOf(lastErr); ra > 0 {
-				backoff = ra
-			}
-			backoff += time.Duration(rand.Intn(250)) * time.Millisecond
-			select {
-			case <-ctx.Done():
-				return nil, ctx.Err()
-			case <-time.After(backoff):
-			}
-		}
-		resp, retryable, err := c.attempt(ctx, payload)
-		if err == nil {
-			return resp, nil
-		}
-		lastErr = err
-		if ctx.Err() != nil {
-			return nil, ctx.Err()
-		}
-		if !retryable {
-			return nil, err
-		}
-		if c.log != nil {
-			c.log.WarnContext(ctx, "anthropic attempt failed, will retry", "attempt", attempt+1, "max", c.profile.MaxAttempts, "err", err)
-		}
-	}
-	return nil, fmt.Errorf("anthropic: exhausted %d attempts: %w", c.profile.MaxAttempts, lastErr)
+	return retryLoop(ctx, c.profile, "anthropic", c.log, func() (*LLMResponse, bool, error) {
+		return c.attempt(ctx, payload)
+	})
 }
 
 // buildRequest maps the neutral request onto the Messages wire. Leading
 // system-role messages become system blocks; CacheBoundary flags become
-// cache_control on that block. Thinking stays off (Фаза 0: Anthropic — только
-// SFW-роли редактора/эскалации, thinking-глубина не нужна и биллится как
-// output; ReasoningEffort пока не маппится — задокументированное ограничение).
+// cache_control on that block. Thinking is not mapped yet (Фаза 0: Anthropic —
+// только SFW-роли редактора/эскалации): ""/"off" проходят (без thinking-
+// конфигурации), а low/medium/high падают громко — молча оплаченная стадия БЕЗ
+// заказанной глубины рассуждений хуже ошибки конфига.
 func (c *anthropicClient) buildRequest(req LLMRequest) (*anthropicRequest, error) {
+	switch req.ReasoningEffort {
+	case "", "off":
+	default:
+		return nil, fmt.Errorf("anthropic: reasoning %q is not mapped by this adapter yet (thinking — Фаза 2)", req.ReasoningEffort)
+	}
 	wire := &anthropicRequest{
-		Model:       req.Model,
-		MaxTokens:   req.MaxTokens,
-		Temperature: req.Temperature,
+		Model:     req.Model,
+		MaxTokens: req.MaxTokens,
 	}
 	boundaries := 0
 	block := func(m Message) anthropicTextBlock {
@@ -242,7 +217,8 @@ func (c *anthropicClient) attempt(ctx context.Context, payload []byte) (*LLMResp
 
 	var out anthropicResponse
 	if err := json.Unmarshal(data, &out); err != nil {
-		return nil, false, fmt.Errorf("anthropic decode: %w", err)
+		// 2xx уже оплачен провайдером — типизируем, retryable (см. httpllm.go).
+		return nil, true, &BilledDecodeError{Provider: "anthropic", Err: err}
 	}
 
 	var text string
diff --git a/backend/internal/llm/provider_anthropic_test.go b/backend/internal/llm/provider_anthropic_test.go
index 64631a1..b466469 100644
--- a/backend/internal/llm/provider_anthropic_test.go
+++ b/backend/internal/llm/provider_anthropic_test.go
@@ -68,6 +68,21 @@ func TestAnthropicWireShapeAndUsage(t *testing.T) {
 	if resp.Text != "перевод" || resp.FinishReason != FinishStop || resp.Model != "claude-sonnet-5" {
 		t.Fatalf("resp = %+v", resp)
 	}
+	// Sampling-параметры на этот wire не отправляются: актуальные Claude-модели
+	// отклоняют temperature с 400.
+	if _, has := got["temperature"]; has {
+		t.Fatal("temperature must not be sent to the Anthropic wire")
+	}
+}
+
+func TestAnthropicRejectsUnmappedReasoning(t *testing.T) {
+	c := NewAnthropicClient(AnthropicConfig{BaseURL: "http://unused", APIKey: "k", Profile: fastProfile()}, nil)
+	_, err := c.Complete(context.Background(), LLMRequest{
+		Model: "m", Messages: []Message{{Role: "user", Content: "u"}}, MaxTokens: 16, ReasoningEffort: "high",
+	})
+	if err == nil {
+		t.Fatal("unmapped reasoning depth must fail loud, not silently run without thinking")
+	}
 }
 
 func TestAnthropicStopReasonMapping(t *testing.T) {
diff --git a/backend/internal/llm/provider_local.go b/backend/internal/llm/provider_local.go
index 87c76d7..e11ff8c 100644
--- a/backend/internal/llm/provider_local.go
+++ b/backend/internal/llm/provider_local.go
@@ -63,10 +63,6 @@ func NewLocalClient(cfg LocalConfig, logger *slog.Logger) LLMClient {
 }
 
 func (c *localClient) Complete(ctx context.Context, req LLMRequest) (*LLMResponse, error) {
-	msgs := make([]openAIMessage, len(req.Messages))
-	for i, m := range req.Messages {
-		msgs[i] = openAIMessage{Role: m.Role, Content: m.Content}
-	}
 	maxTok := req.MaxTokens
 	if c.maxTok > 0 {
 		maxTok = c.maxTok
@@ -75,10 +71,6 @@ func (c *localClient) Complete(ctx context.Context, req LLMRequest) (*LLMRespons
 	if c.temp > 0 {
 		temp = c.temp
 	}
-	var respFormat any
-	if req.JSONOnly {
-		respFormat = map[string]string{"type": "json_object"}
-	}
 	effort := req.ReasoningEffort
 	if effort == "off" {
 		// ollama's /v1 maps reasoning_effort onto Qwen3 thinking; "none" is its
@@ -87,12 +79,12 @@ func (c *localClient) Complete(ctx context.Context, req LLMRequest) (*LLMRespons
 	}
 	resp, err := c.http.complete(ctx, openAIRequest{
 		Model:           c.model,
-		Messages:        msgs,
+		Messages:        toOpenAIMessages(req.Messages),
 		MaxTokens:       maxTok,
 		Temperature:     temp,
 		Stream:          false,
 		ReasoningEffort: effort,
-		ResponseFormat:  respFormat,
+		ResponseFormat:  jsonResponseFormat(req.JSONOnly),
 	})
 	if err != nil {
 		return nil, err
diff --git a/backend/internal/llm/provider_openai.go b/backend/internal/llm/provider_openai.go
index cf6e07a..c611eb6 100644
--- a/backend/internal/llm/provider_openai.go
+++ b/backend/internal/llm/provider_openai.go
@@ -64,18 +64,6 @@ func NewOpenAICompatClient(cfg OpenAICompatConfig, logger *slog.Logger) LLMClien
 }
 
 func (c *openAICompatClient) Complete(ctx context.Context, req LLMRequest) (*LLMResponse, error) {
-	msgs := make([]openAIMessage, len(req.Messages))
-	for i, m := range req.Messages {
-		// CacheBoundary is meaningless on this wire: OpenAI-compatible caches
-		// (DeepSeek: prefix-match с 0-го токена блоками по 64 токена) are
-		// automatic. The stable-prefix ordering the assembler produced is all
-		// that matters here.
-		msgs[i] = openAIMessage{Role: m.Role, Content: m.Content}
-	}
-	var respFormat any
-	if req.JSONOnly {
-		respFormat = map[string]string{"type": "json_object"}
-	}
 	effort := req.ReasoningEffort
 	if effort == "off" {
 		// "off" is our neutral value; OpenAI-compat providers with a
@@ -84,12 +72,12 @@ func (c *openAICompatClient) Complete(ctx context.Context, req LLMRequest) (*LLM
 	}
 	resp, err := c.http.complete(ctx, openAIRequest{
 		Model:           req.Model,
-		Messages:        msgs,
+		Messages:        toOpenAIMessages(req.Messages),
 		MaxTokens:       req.MaxTokens,
 		Temperature:     req.Temperature,
 		Stream:          false,
 		ReasoningEffort: effort,
-		ResponseFormat:  respFormat,
+		ResponseFormat:  jsonResponseFormat(req.JSONOnly),
 		extra:           c.extra,
 	})
 	if err != nil {
diff --git a/backend/internal/obs/logging.go b/backend/internal/obs/logging.go
index e408b27..8bebd85 100644
--- a/backend/internal/obs/logging.go
+++ b/backend/internal/obs/logging.go
@@ -30,15 +30,23 @@ func NewLogger() *slog.Logger {
 }
 
 // contextHandler wraps a slog.Handler so a record logged with one of the
-// *Context methods automatically gets the call's trace_id attached — the
-// userver-style trail, set once at the top of a unit of work and carried
-// through to the model call. Records logged without a traced context simply
-// carry no trace_id.
+// *Context methods automatically gets the call's trace_id AND the translation
+// axis (book/chapter/chunk/stage/role) attached — the userver-style trail, set
+// once at the top of a unit of work and carried through to the model call.
+// Records logged without a traced context simply carry none of these.
 type contextHandler struct{ slog.Handler }
 
 func (h contextHandler) Handle(ctx context.Context, r slog.Record) error {
-	if id := TraceFromContext(ctx); id != "" {
-		r.AddAttrs(slog.String("trace_id", id))
+	if ri, ok := ReqInfoFromContext(ctx); ok {
+		if ri.TraceID != "" {
+			r.AddAttrs(slog.String("trace_id", ri.TraceID))
+		}
+		if ri.Book != "" {
+			r.AddAttrs(slog.String("book", ri.Book), slog.Int("chapter", ri.Chapter), slog.Int("chunk", ri.Chunk))
+		}
+		if ri.Stage != "" {
+			r.AddAttrs(slog.String("stage", ri.Stage), slog.String("role", ri.Role))
+		}
 	}
 	return h.Handler.Handle(ctx, r)
 }
diff --git a/backend/internal/pipeline/clients.go b/backend/internal/pipeline/clients.go
index 656677e..63362bc 100644
--- a/backend/internal/pipeline/clients.go
+++ b/backend/internal/pipeline/clients.go
@@ -38,11 +38,12 @@ func BuildClient(models *config.Models, modelName string, logger *slog.Logger) (
 		}, logger), nil
 	case "local":
 		return llm.NewLocalClient(llm.LocalConfig{
-			BaseURL:   prov.BaseURL,
-			APIKey:    prov.APIKey(),
-			Model:     prov.Model,
-			MaxTokens: prov.MaxTokens,
-			Profile:   prov.Timeouts.Profile(),
+			BaseURL:     prov.BaseURL,
+			APIKey:      prov.APIKey(),
+			Model:       prov.Model,
+			Temperature: prov.Temperature,
+			MaxTokens:   prov.MaxTokens,
+			Profile:     prov.Timeouts.Profile(),
 		}, logger), nil
 	default:
 		return nil, fmt.Errorf("pipeline: provider %q has unknown kind %q", mod.Provider, prov.Kind)
diff --git a/backend/internal/pipeline/render.go b/backend/internal/pipeline/render.go
index 896a632..a596c3a 100644
--- a/backend/internal/pipeline/render.go
+++ b/backend/internal/pipeline/render.go
@@ -30,6 +30,12 @@ import (
 // (§3.4 — [НУЖНО РЕШЕНИЕ] п.2в).
 const chunkerVersion = "chunker-v0-wholefile"
 
+// estimatorVersion versions EstimateTokens: его выход входит в max_tokens и
+// через него в request-hash, поэтому перекалибровка весов — тоже явная
+// инвалидация через snapshot, а не тихий промах всех чекпоинтов (находка
+// ревью: code-only правка эстиматора пере-оплатила бы полкниги).
+const estimatorVersion = "estimator-v0"
+
 // userSeparator splits a prompt template file into the system part and the
 // user part. Обе части — версионируемые шаблоны в prompts/ («Редакция» позже
 // правит файлы, не код).
@@ -67,37 +73,57 @@ type RenderVars struct {
 	Draft string // prior stage output ("" on the first stage)
 }
 
-// Render substitutes placeholders. Unknown {{…}} markers are a hard error —
-// молчаливо пустой плейсхолдер в промпте хуже падения.
+// Render substitutes placeholders in a SINGLE pass over the template: values
+// are inserted verbatim and never re-scanned, so literal «{{…}}» внутри
+// исходника или черновика LLM — это просто текст, а не маркер (находка ревью:
+// последовательный ReplaceAll подставлял черновик в «{{draft}}» из сорса и
+// падал на «{{TN: примечание}}» в выводе модели). Unknown {{…}} markers in
+// the TEMPLATE are a hard error — молчаливо пустой плейсхолдер хуже падения.
 func Render(tpl string, v RenderVars) (string, error) {
-	pairs := [...][2]string{
-		{"{{book_id}}", v.Book.BookID},
-		{"{{title}}", v.Book.Title},
-		{"{{source_lang}}", v.Book.SourceLang},
-		{"{{target_lang}}", v.Book.TargetLang},
-		{"{{genre}}", v.Book.Genre},
-		{"{{audience}}", v.Book.Audience},
-		{"{{venuti}}", strconv.FormatFloat(v.Book.Venuti, 'f', 2, 64)},
-		{"{{honorifics}}", v.Book.Honorifics},
-		{"{{transcription}}", v.Book.Transcription},
-		{"{{footnotes}}", v.Book.Footnotes},
-		{"{{text}}", v.Text},
-		{"{{draft}}", v.Draft},
+	// Lookup-only map (никакой итерации — детерминизм не страдает).
+	vals := map[string]string{
+		"book_id":       v.Book.BookID,
+		"title":         v.Book.Title,
+		"source_lang":   v.Book.SourceLang,
+		"target_lang":   v.Book.TargetLang,
+		"genre":         v.Book.Genre,
+		"audience":      v.Book.Audience,
+		"venuti":        strconv.FormatFloat(v.Book.Venuti, 'f', 2, 64),
+		"honorifics":    v.Book.Honorifics,
+		"transcription": v.Book.Transcription,
+		"footnotes":     v.Book.Footnotes,
+		"text":          v.Text,
+		"draft":         v.Draft,
 	}
-	out := tpl
-	for _, p := range pairs {
-		out = strings.ReplaceAll(out, p[0], p[1])
-	}
-	if i := strings.Index(out, "{{"); i >= 0 {
-		end := strings.Index(out[i:], "}}")
-		if end < 0 {
-			end = len(out) - i
-		} else {
-			end += 2
+	var b strings.Builder
+	rest := tpl
+	for {
+		i := strings.Index(rest, "{{")
+		if i < 0 {
+			b.WriteString(rest)
+			return b.String(), nil
 		}
-		return "", fmt.Errorf("pipeline: unknown placeholder %q in template", out[i:i+end])
+		b.WriteString(rest[:i])
+		rest = rest[i:]
+		end := strings.Index(rest, "}}")
+		if end < 0 {
+			return "", fmt.Errorf("pipeline: unterminated placeholder %q in template", snippetStr(rest))
+		}
+		name := rest[2:end]
+		val, ok := vals[name]
+		if !ok {
+			return "", fmt.Errorf("pipeline: unknown placeholder {{%s}} in template", name)
+		}
+		b.WriteString(val)
+		rest = rest[end+2:]
 	}
-	return out, nil
+}
+
+func snippetStr(s string) string {
+	if len(s) > 40 {
+		return s[:40] + "…"
+	}
+	return s
 }
 
 // Messages builds the wire-neutral message list for a stage. Layout по Р5:
@@ -122,8 +148,10 @@ func Messages(tpl *PromptTemplate, v RenderVars) ([]llm.Message, error) {
 // RequestHash is the checkpoint/resume key (§3.1): a stable hash of everything
 // that determines the call. The snapshot id freezes the volatile context, so
 // identical re-renders after a restart find their checkpoint and are neither
-// repeated nor re-billed.
-func RequestHash(bookID string, chapter, chunkIdx int, stage, role, model string, temperature float64, reasoning string, jsonOnly bool, maxTokens int, snapshotID string, msgs []llm.Message) string {
+// repeated nor re-billed. attempt — измерение регенераций (Фаза 1: гейт
+// провалился → attempt+1 → новый ключ; без него регенерация попадала бы в
+// только что провалившийся чекпоинт и мгновенно эскалировала в премиум).
+func RequestHash(bookID string, chapter, chunkIdx, attempt int, stage, role, model string, temperature float64, reasoning string, jsonOnly bool, maxTokens int, snapshotID string, msgs []llm.Message) string {
 	h := sha256.New()
 	w := func(parts ...string) {
 		for _, p := range parts {
@@ -131,7 +159,7 @@ func RequestHash(bookID string, chapter, chunkIdx int, stage, role, model string
 			h.Write([]byte{0})
 		}
 	}
-	w("tm-request-v1", bookID, strconv.Itoa(chapter), strconv.Itoa(chunkIdx), stage, role, model,
+	w("tm-request-v1", bookID, strconv.Itoa(chapter), strconv.Itoa(chunkIdx), strconv.Itoa(attempt), stage, role, model,
 		strconv.FormatFloat(temperature, 'f', -1, 64), reasoning, strconv.FormatBool(jsonOnly),
 		strconv.Itoa(maxTokens), snapshotID)
 	for _, m := range msgs {
diff --git a/backend/internal/pipeline/render_test.go b/backend/internal/pipeline/render_test.go
index a93d40c..919e863 100644
--- a/backend/internal/pipeline/render_test.go
+++ b/backend/internal/pipeline/render_test.go
@@ -3,6 +3,7 @@ package pipeline
 import (
 	"os"
 	"path/filepath"
+	"strings"
 	"testing"
 
 	"textmachine/backend/internal/config"
@@ -55,8 +56,8 @@ func TestRenderIsDeterministic(t *testing.T) {
 			}
 		}
 	}
-	h1 := RequestHash("b", 1, 0, "draft", "translator", "m", 0.3, "off", false, 1024, "snap", m1)
-	h2 := RequestHash("b", 1, 0, "draft", "translator", "m", 0.3, "off", false, 1024, "snap", m1)
+	h1 := RequestHash("b", 1, 0, 0, "draft", "translator", "m", 0.3, "off", false, 1024, "snap", m1)
+	h2 := RequestHash("b", 1, 0, 0, "draft", "translator", "m", 0.3, "off", false, 1024, "snap", m1)
 	if h1 != h2 {
 		t.Fatal("request hash is not stable")
 	}
@@ -64,15 +65,15 @@ func TestRenderIsDeterministic(t *testing.T) {
 
 func TestRequestHashSensitivity(t *testing.T) {
 	msgs := []llm.Message{{Role: "system", Content: "s", CacheBoundary: true}, {Role: "user", Content: "u"}}
-	base := RequestHash("b", 1, 0, "draft", "translator", "m", 0.3, "off", false, 1024, "snap", msgs)
+	base := RequestHash("b", 1, 0, 0, "draft", "translator", "m", 0.3, "off", false, 1024, "snap", msgs)
 
 	variants := map[string]string{
-		"model":     RequestHash("b", 1, 0, "draft", "translator", "m2", 0.3, "off", false, 1024, "snap", msgs),
-		"temp":      RequestHash("b", 1, 0, "draft", "translator", "m", 0.4, "off", false, 1024, "snap", msgs),
-		"snapshot":  RequestHash("b", 1, 0, "draft", "translator", "m", 0.3, "off", false, 1024, "snap2", msgs),
-		"maxTokens": RequestHash("b", 1, 0, "draft", "translator", "m", 0.3, "off", false, 2048, "snap", msgs),
-		"chunk":     RequestHash("b", 1, 1, "draft", "translator", "m", 0.3, "off", false, 1024, "snap", msgs),
-		"content":   RequestHash("b", 1, 0, "draft", "translator", "m", 0.3, "off", false, 1024, "snap", []llm.Message{{Role: "system", Content: "s", CacheBoundary: true}, {Role: "user", Content: "u2"}}),
+		"model":     RequestHash("b", 1, 0, 0, "draft", "translator", "m2", 0.3, "off", false, 1024, "snap", msgs),
+		"temp":      RequestHash("b", 1, 0, 0, "draft", "translator", "m", 0.4, "off", false, 1024, "snap", msgs),
+		"snapshot":  RequestHash("b", 1, 0, 0, "draft", "translator", "m", 0.3, "off", false, 1024, "snap2", msgs),
+		"maxTokens": RequestHash("b", 1, 0, 0, "draft", "translator", "m", 0.3, "off", false, 2048, "snap", msgs),
+		"chunk":     RequestHash("b", 1, 1, 0, "draft", "translator", "m", 0.3, "off", false, 1024, "snap", msgs),
+		"content":   RequestHash("b", 1, 0, 0, "draft", "translator", "m", 0.3, "off", false, 1024, "snap", []llm.Message{{Role: "system", Content: "s", CacheBoundary: true}, {Role: "user", Content: "u2"}}),
 	}
 	for name, h := range variants {
 		if h == base {
@@ -81,9 +82,9 @@ func TestRequestHashSensitivity(t *testing.T) {
 	}
 
 	// Конкатенация с разделителем не склеивается: ("ab","c") != ("a","bc").
-	a := RequestHash("b", 1, 0, "draft", "translator", "m", 0.3, "off", false, 1024, "snap",
+	a := RequestHash("b", 1, 0, 0, "draft", "translator", "m", 0.3, "off", false, 1024, "snap",
 		[]llm.Message{{Role: "user", Content: "ab"}, {Role: "user", Content: "c"}})
-	b := RequestHash("b", 1, 0, "draft", "translator", "m", 0.3, "off", false, 1024, "snap",
+	b := RequestHash("b", 1, 0, 0, "draft", "translator", "m", 0.3, "off", false, 1024, "snap",
 		[]llm.Message{{Role: "user", Content: "a"}, {Role: "user", Content: "bc"}})
 	if a == b {
 		t.Fatal("field separator is not injective")
@@ -98,6 +99,27 @@ func TestRenderRejectsUnknownPlaceholder(t *testing.T) {
 	}
 }
 
+// Литеральные «{{…}}» в ИСХОДНИКЕ/черновике — просто текст: значения
+// вставляются verbatim и не пере-сканируются (однопроходный рендер).
+func TestRenderValuesAreNotRescanned(t *testing.T) {
+	tpl := writeTemplate(t, "Система.\n---USER---\nТекст: {{text}}\nЧерновик: {{draft}}")
+	v := RenderVars{
+		Book:  testBook(),
+		Text:  "герой сказал {{draft}} и ушёл", // literal braces in source
+		Draft: "примечание {{TN: сноска}} осталось", // literal braces in prior LLM output
+	}
+	msgs, err := Messages(tpl, v)
+	if err != nil {
+		t.Fatalf("literal braces inside values must not fail: %v", err)
+	}
+	user := msgs[1].Content
+	for _, want := range []string{"герой сказал {{draft}} и ушёл", "примечание {{TN: сноска}} осталось"} {
+		if !strings.Contains(user, want) {
+			t.Fatalf("value %q must survive verbatim, got:\n%s", want, user)
+		}
+	}
+}
+
 func TestTemplateRequiresUserSeparator(t *testing.T) {
 	path := filepath.Join(t.TempDir(), "bad.md")
 	if err := os.WriteFile(path, []byte("только системная часть"), 0o644); err != nil {
diff --git a/backend/internal/pipeline/runner.go b/backend/internal/pipeline/runner.go
index e258aad..befa92c 100644
--- a/backend/internal/pipeline/runner.go
+++ b/backend/internal/pipeline/runner.go
@@ -5,6 +5,7 @@ import (
 	"crypto/sha256"
 	"encoding/hex"
 	"encoding/json"
+	"errors"
 	"fmt"
 	"log/slog"
 	"os"
@@ -32,6 +33,11 @@ type Runner struct {
 	Store    *store.Store
 	Pricer   *ledger.Pricer
 	Log      *slog.Logger
+	// Resnapshot re-pins existing jobs to the CURRENT snapshot when configs/
+	// prompts changed since the job started (tmctl --resnapshot). Без флага
+	// расхождение — громкая ошибка: смена контекста инвалидирует чекпоинты и
+	// пере-оплачивает вызовы, это делается только осознанно (Р6).
+	Resnapshot bool
 
 	clients   map[string]llm.LLMClient
 	templates map[string]*PromptTemplate
@@ -117,14 +123,16 @@ func (r *Runner) snapshotID() (id, payload string, err error) {
 		Reasoning     string  `json:"reasoning"`
 	}
 	snap := struct {
-		BriefHash      string      `json:"brief_hash"`
-		ChunkerVersion string      `json:"chunker_version"`
-		PipelineCore   string      `json:"pipeline_core"`
-		Stages         []stageSnap `json:"stages"`
+		BriefHash        string      `json:"brief_hash"`
+		ChunkerVersion   string      `json:"chunker_version"`
+		EstimatorVersion string      `json:"estimator_version"`
+		PipelineCore     string      `json:"pipeline_core"`
+		Stages           []stageSnap `json:"stages"`
 	}{
-		BriefHash:      r.Book.BriefHash(),
-		ChunkerVersion: chunkerVersion,
-		PipelineCore:   r.Pipeline.Core,
+		BriefHash:        r.Book.BriefHash(),
+		ChunkerVersion:   chunkerVersion,
+		EstimatorVersion: estimatorVersion,
+		PipelineCore:     r.Pipeline.Core,
 	}
 	for _, st := range r.Pipeline.Stages {
 		snap.Stages = append(snap.Stages, stageSnap{
@@ -207,6 +215,19 @@ func (r *Runner) runStage(ctx context.Context, st config.Stage, snapID string, c
 	if err != nil {
 		return nil, err
 	}
+	// Snapshot pinning (Р6): джоба заморожена на снапшоте своего старта. Если
+	// конфиги/промпты с тех пор изменились — это инвалидация чекпоинтов и
+	// пере-оплата; выполняем её только по явной команде.
+	if job.SnapshotID != snapID {
+		if !r.Resnapshot {
+			return nil, fmt.Errorf("pipeline: job %s/ch%d/%s was started under snapshot %.12s, current config renders snapshot %.12s — конфиг/промпты изменились; уже оплаченные чекпоинты станут недействительны и вызовы будут пере-оплачены; повторить с --resnapshot, чтобы принять это явно",
+				r.Book.BookID, chapter, st.Name, job.SnapshotID, snapID)
+		}
+		if err := r.Store.UpdateJobSnapshot(job.ID, snapID); err != nil {
+			return nil, err
+		}
+		r.Log.WarnContext(ctx, "job re-pinned to new snapshot (--resnapshot)", "stage", st.Name, "old", job.SnapshotID[:12], "new", snapID[:12])
+	}
 	if err := r.Store.SetJobStatus(job.ID, "running"); err != nil {
 		return nil, err
 	}
@@ -225,19 +246,27 @@ func (r *Runner) runStage(ctx context.Context, st config.Stage, snapID string, c
 		maxTokens = r.Pipeline.Defaults.MinMaxTokens
 	}
 
-	reqHash := RequestHash(r.Book.BookID, chapter, chunkIdx, st.Name, st.Role, st.Model,
+	// attempt 0: измерение регенераций подключит гейт-цикл Фазы 1.
+	const attempt = 0
+	reqHash := RequestHash(r.Book.BookID, chapter, chunkIdx, attempt, st.Name, st.Role, st.Model,
 		st.Temperature, st.Reasoning, false, maxTokens, snapID, msgs)
 
-	ctx = obs.WithReqInfo(ctx, obs.ReqInfo{
-		TraceID: obs.TraceFromContext(ctx), Book: r.Book.BookID,
-		Chapter: chapter, Chunk: chunkIdx, Stage: st.Name, Role: st.Role,
-	})
+	// Дополняем ReqInfo, СОХРАНЯЯ решения допуска (LogBodies) — перезапись с
+	// нуля отрезала бы задокументированный debug-канал (находка ревью).
+	ri, _ := obs.ReqInfoFromContext(ctx)
+	ri.Book, ri.Chapter, ri.Chunk, ri.Stage, ri.Role = r.Book.BookID, chapter, chunkIdx, st.Name, st.Role
+	ctx = obs.WithReqInfo(ctx, ri)
 
 	// Resume path: a checkpoint means the call already happened and was paid —
 	// serve it, log the hit, never re-bill (приёмка: kill -9 теряет ≤1 вызов).
+	// Пустой чекпоинт (оплаченный отказ/обрезка/нечитаемое тело) НЕ подаётся
+	// дальше молча: черновик="" отравил бы редактуру и экспорт (находка ревью).
 	if cp, err := r.Store.GetCheckpoint(reqHash); err != nil {
 		return nil, err
 	} else if cp != nil {
+		if strings.TrimSpace(cp.ResponseText) == "" {
+			return nil, fmt.Errorf("pipeline: stage %s has a BILLED but empty checkpoint (finish=%s, hash %.12s) — вызов был оплачен, но пригодного текста нет; до регенераций Фазы 1: удалить строку checkpoints вручную или дождаться механизма attempt", st.Name, cp.FinishReason, reqHash)
+		}
 		var usage llm.Usage
 		_ = json.Unmarshal([]byte(cp.UsageJSON), &usage)
 		r.Store.LogRequest(ctx, r.Log, store.RequestLog{
@@ -289,6 +318,27 @@ func (r *Runner) runStage(ctx context.Context, st config.Stage, snapID string, c
 	})
 	latency := int(time.Since(start).Milliseconds())
 	if err != nil {
+		// 2xx с нечитаемым телом: провайдер УЖЕ списал деньги — резерв не
+		// возвращаем, консервативно селтлим ОЦЕНКУ с пустым чекпоинтом
+		// (переучёт ≤1 оценки безопаснее слепого потолка) и падаем громко.
+		var bde *llm.BilledDecodeError
+		if errors.As(err, &bde) {
+			if serr := r.Store.SettleWithCheckpoint(resv, estimate, store.Checkpoint{
+				RequestHash: reqHash, JobID: job.ID, ChunkIdx: chunkIdx, Attempt: attempt,
+				Stage: st.Name, Role: st.Role, ModelRequested: st.Model, ModelActual: st.Model,
+				ResponseText: "", UsageJSON: "{}", CostUSD: estimate, FinishReason: "decode_error",
+			}); serr != nil {
+				r.Log.ErrorContext(ctx, "settle after billed decode failure also failed", "err", serr)
+			}
+			r.Store.LogRequest(ctx, r.Log, store.RequestLog{
+				BookID: r.Book.BookID, Chapter: chapter, ChunkIdx: chunkIdx,
+				Stage: st.Name, Role: st.Role, ModelRequested: st.Model,
+				RequestHash: reqHash, CostUSD: estimate, LatencyMS: latency,
+				Degraded: "billed_2xx_decode_failed", Err: err.Error(), OK: false,
+			})
+			_ = r.Store.SetJobStatus(job.ID, "failed")
+			return nil, fmt.Errorf("pipeline: stage %s: %w (оценка $%.6f заселтлена консервативно)", st.Name, err, estimate)
+		}
 		// No 2xx ever arrived: nothing was billed by the provider — release the
 		// reservation and surface the failure with a telemetry row.
 		if rerr := r.Store.Release(resv); rerr != nil {
@@ -315,7 +365,8 @@ func (r *Runner) runStage(ctx context.Context, st config.Stage, snapID string, c
 
 	// Деньги: settle + сырой ответ — одна транзакция (§3.3).
 	if err := r.Store.SettleWithCheckpoint(resv, cost, store.Checkpoint{
-		RequestHash: reqHash, JobID: job.ID, ChunkIdx: chunkIdx, Stage: st.Name, Role: st.Role,
+		RequestHash: reqHash, JobID: job.ID, ChunkIdx: chunkIdx, Attempt: attempt,
+		Stage: st.Name, Role: st.Role,
 		ModelRequested: st.Model, ModelActual: modelActual, ResponseText: resp.Text,
 		UsageJSON: string(usageJSON), CostUSD: cost, FinishReason: resp.FinishReason,
 		ProviderRequestID: resp.ProviderRequestID,
@@ -325,6 +376,25 @@ func (r *Runner) runStage(ctx context.Context, st config.Stage, snapID string, c
 		return nil, fmt.Errorf("pipeline: settle stage %s: %w", st.Name, err)
 	}
 
+	// Пустой оплаченный ответ (finish=length: thinking съел бюджет; фильтр):
+	// деньги заселтлены, чекпоинт есть — но дальше по конвейеру его НЕ подаём:
+	// черновик="" молча отравил бы редактуру. Громкая остановка (регенерация
+	// по attempt — Фаза 1).
+	if strings.TrimSpace(resp.Text) == "" {
+		r.Store.LogRequest(ctx, r.Log, store.RequestLog{
+			BookID: r.Book.BookID, Chapter: chapter, ChunkIdx: chunkIdx,
+			Stage: st.Name, Role: st.Role, ModelRequested: st.Model, ModelActual: modelActual,
+			RequestHash:  reqHash,
+			PromptTokens: resp.Usage.PromptTokens, CachedTokens: resp.Usage.CachedTokens,
+			CacheCreationTokens: resp.Usage.CacheCreationTokens,
+			CompletionTokens:    resp.Usage.CompletionTokens, ReasoningTokens: resp.Usage.ReasoningTokens,
+			CostUSD: cost, LatencyMS: latency, FinishReason: resp.FinishReason,
+			Degraded: "empty_completion", OK: false,
+		})
+		_ = r.Store.SetJobStatus(job.ID, "failed")
+		return nil, fmt.Errorf("pipeline: stage %s returned an EMPTY completion (finish=%s, billed $%.6f, checkpointed) — вероятно thinking/фильтр съел бюджет; поднимите max_tokens/смените модель и удалите строку checkpoints %.12s", st.Name, resp.FinishReason, cost, reqHash)
+	}
+
 	r.Store.LogRequest(ctx, r.Log, store.RequestLog{
 		BookID: r.Book.BookID, Chapter: chapter, ChunkIdx: chunkIdx,
 		Stage: st.Name, Role: st.Role, ModelRequested: st.Model, ModelActual: modelActual,
diff --git a/backend/internal/pipeline/runner_test.go b/backend/internal/pipeline/runner_test.go
index b1c0bef..c081fe3 100644
--- a/backend/internal/pipeline/runner_test.go
+++ b/backend/internal/pipeline/runner_test.go
@@ -168,6 +168,63 @@ func TestRunnerEndToEndWithResume(t *testing.T) {
 	}
 }
 
+// Р6: правка конфига/промпта после старта джоб — не тихая инвалидация
+// чекпоинтов, а громкая ошибка; пере-перевод только явным --resnapshot.
+func TestRunnerSnapshotPinning(t *testing.T) {
+	var calls atomic.Int32
+	srv := newFakeProvider(t, &calls)
+	defer srv.Close()
+	bookPath := setupProject(t, srv.URL)
+	ctx := context.Background()
+
+	r1, err := NewRunner(bookPath, obs.NewLogger())
+	if err != nil {
+		t.Fatal(err)
+	}
+	if _, err := r1.TranslateOneChunk(ctx); err != nil {
+		t.Fatal(err)
+	}
+	r1.Close()
+	if calls.Load() != 2 {
+		t.Fatalf("run1 calls = %d", calls.Load())
+	}
+
+	// Меняем промпт переводчика → новый snapshot.
+	promptPath := filepath.Join(filepath.Dir(bookPath), "prompts", "translator.md")
+	writeFile(t, promptPath, "НОВЫЙ промпт с {{source_lang}}.\n---USER---\n{{text}}")
+
+	r2, err := NewRunner(bookPath, obs.NewLogger())
+	if err != nil {
+		t.Fatal(err)
+	}
+	_, err = r2.TranslateOneChunk(ctx)
+	r2.Close()
+	if err == nil || !strings.Contains(err.Error(), "resnapshot") {
+		t.Fatalf("changed config must fail loud mentioning --resnapshot, got: %v", err)
+	}
+	if calls.Load() != 2 {
+		t.Fatalf("denied resume must not call the provider, calls=%d", calls.Load())
+	}
+
+	// Явное согласие: джобы перепривязываются, вызовы повторяются и оплачиваются.
+	r3, err := NewRunner(bookPath, obs.NewLogger())
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer r3.Close()
+	r3.Resnapshot = true
+	res, err := r3.TranslateOneChunk(ctx)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if calls.Load() != 4 {
+		t.Fatalf("resnapshot run must re-call both stages, calls=%d", calls.Load())
+	}
+	if res.TotalUSD <= 0 {
+		t.Fatal("re-translation must be billed")
+	}
+}
+
 func TestRunnerCeilingDenies(t *testing.T) {
 	var calls atomic.Int32
 	srv := newFakeProvider(t, &calls)
diff --git a/backend/internal/store/jobs.go b/backend/internal/store/jobs.go
index c9468a1..5cbe796 100644
--- a/backend/internal/store/jobs.go
+++ b/backend/internal/store/jobs.go
@@ -71,3 +71,14 @@ func (s *Store) SetJobStatus(jobID int64, status string) error {
 		`UPDATE jobs SET status = ?, updated_at = datetime('now') WHERE id = ?`, status, jobID)
 	return err
 }
+
+// UpdateJobSnapshot re-pins a job to a new snapshot — ЯВНАЯ команда оператора
+// (tmctl --resnapshot): пере-перевод уже оплаченных чанков случается только
+// осознанно, никогда как тихий побочный эффект правки конфига (Р6).
+func (s *Store) UpdateJobSnapshot(jobID int64, snapshotID string) error {
+	ctx, cancel := opContext()
+	defer cancel()
+	_, err := s.w.ExecContext(ctx,
+		`UPDATE jobs SET snapshot_id = ?, updated_at = datetime('now') WHERE id = ?`, snapshotID, jobID)
+	return err
+}
diff --git a/backend/internal/store/ledger.go b/backend/internal/store/ledger.go
index e7edc4d..dbadeaf 100644
--- a/backend/internal/store/ledger.go
+++ b/backend/internal/store/ledger.go
@@ -99,6 +99,7 @@ type Checkpoint struct {
 	RequestHash       string
 	JobID             int64
 	ChunkIdx          int
+	Attempt           int
 	Stage             string
 	Role              string
 	ModelRequested    string
@@ -132,12 +133,12 @@ func (s *Store) SettleWithCheckpoint(res Reservation, cost float64, cp Checkpoin
 
 	ins, err := tx.ExecContext(ctx, `
 		INSERT INTO checkpoints (
-			request_hash, job_id, chunk_idx, stage, role,
+			request_hash, job_id, chunk_idx, attempt, stage, role,
 			model_requested, model_actual, response_text, usage_json,
 			cost_usd, finish_reason, provider_request_id
-		) VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
+		) VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
 		ON CONFLICT (request_hash) DO NOTHING`,
-		cp.RequestHash, cp.JobID, cp.ChunkIdx, cp.Stage, cp.Role,
+		cp.RequestHash, cp.JobID, cp.ChunkIdx, cp.Attempt, cp.Stage, cp.Role,
 		cp.ModelRequested, cp.ModelActual, cp.ResponseText, cp.UsageJSON,
 		cp.CostUSD, cp.FinishReason, cp.ProviderRequestID)
 	if err != nil {
@@ -173,10 +174,10 @@ func (s *Store) GetCheckpoint(requestHash string) (*Checkpoint, error) {
 	defer cancel()
 	cp := Checkpoint{RequestHash: requestHash}
 	err := s.r.QueryRowContext(ctx, `
-		SELECT job_id, chunk_idx, stage, role, model_requested, model_actual,
+		SELECT job_id, chunk_idx, attempt, stage, role, model_requested, model_actual,
 		       response_text, usage_json, cost_usd, finish_reason, provider_request_id
 		FROM checkpoints WHERE request_hash = ?`, requestHash).Scan(
-		&cp.JobID, &cp.ChunkIdx, &cp.Stage, &cp.Role, &cp.ModelRequested, &cp.ModelActual,
+		&cp.JobID, &cp.ChunkIdx, &cp.Attempt, &cp.Stage, &cp.Role, &cp.ModelRequested, &cp.ModelActual,
 		&cp.ResponseText, &cp.UsageJSON, &cp.CostUSD, &cp.FinishReason, &cp.ProviderRequestID)
 	if errors.Is(err, sql.ErrNoRows) {
 		return nil, nil
diff --git a/backend/internal/store/migrate.go b/backend/internal/store/migrate.go
index 3e97ac5..f26a37e 100644
--- a/backend/internal/store/migrate.go
+++ b/backend/internal/store/migrate.go
@@ -46,6 +46,7 @@ var migrations = []string{
 		request_hash        TEXT PRIMARY KEY,
 		job_id              INTEGER NOT NULL REFERENCES jobs(id),
 		chunk_idx           INTEGER NOT NULL,
+		attempt             INTEGER NOT NULL DEFAULT 0, -- регенерации Фазы 1: attempt входит в request-hash
 		stage               TEXT NOT NULL,
 		role                TEXT NOT NULL,
 		model_requested     TEXT NOT NULL,
diff --git a/backend/internal/store/requestlog.go b/backend/internal/store/requestlog.go
index 813f2ab..13f5119 100644
--- a/backend/internal/store/requestlog.go
+++ b/backend/internal/store/requestlog.go
@@ -55,7 +55,7 @@ func (s *Store) InsertRequestLog(rl RequestLog) error {
 		rl.ModelRequested, rl.ModelActual, rl.RequestHash,
 		rl.PromptTokens, rl.CachedTokens, rl.CacheCreationTokens,
 		rl.CompletionTokens, rl.ReasoningTokens,
-		rl.CostUSD, rl.LatencyMS, rl.FinishReason, boolToInt(rl.TMHit), rl.Degraded, rl.Err, boolToInt(rl.OK))
+		rl.CostUSD, rl.LatencyMS, rl.FinishReason, rl.TMHit, rl.Degraded, rl.Err, rl.OK)
 	return err
 }
 
@@ -80,10 +80,3 @@ func (s *Store) RequestLogRows(bookID string) (*sql.Rows, error) {
 		       cost_usd, latency_ms, finish_reason, tm_hit, ok
 		FROM request_log WHERE book_id = ? ORDER BY id`, bookID)
 }
-
-func boolToInt(b bool) int {
-	if b {
-		return 1
-	}
-	return 0
-}
diff --git a/backend/internal/store/store.go b/backend/internal/store/store.go
index ec76279..74a3f00 100644
--- a/backend/internal/store/store.go
+++ b/backend/internal/store/store.go
@@ -18,6 +18,8 @@ import (
 	"database/sql"
 	"fmt"
 	"net/url"
+	"os"
+	"syscall"
 	"time"
 
 	_ "modernc.org/sqlite"
@@ -30,13 +32,23 @@ const opTimeout = 10 * time.Second
 
 // Store is one project database.
 type Store struct {
-	w *sql.DB // single-connection write pool; all mutations go through it
-	r *sql.DB // read pool
+	w    *sql.DB // single-connection write pool; all mutations go through it
+	r    *sql.DB // read pool
+	lock *os.File
 }
 
 // Open opens (or creates) the project database at path, applies pending
 // migrations and recovers stale reservations from a crashed run.
+//
+// Ownership: Open takes an EXCLUSIVE flock on <path>.lock — «один процесс
+// владеет файлом проекта» здесь принуждается, а не предполагается. Без замка
+// параллельный `tmctl report` во время перевода обнулил бы recovery-проходом
+// живые резервы бегущего процесса и ослепил бы его потолки (находка ревью).
 func Open(path string) (*Store, error) {
+	lock, err := acquireLock(path + ".lock")
+	if err != nil {
+		return nil, err
+	}
 	dsn := func(txlock string) string {
 		v := url.Values{}
 		v.Add("_pragma", "journal_mode(WAL)")
@@ -51,6 +63,7 @@ func Open(path string) (*Store, error) {
 
 	w, err := sql.Open("sqlite", dsn("immediate"))
 	if err != nil {
+		releaseLock(lock)
 		return nil, fmt.Errorf("store: open write pool: %w", err)
 	}
 	// One writer connection: SQLite has a single writer per file anyway;
@@ -60,11 +73,12 @@ func Open(path string) (*Store, error) {
 	r, err := sql.Open("sqlite", dsn(""))
 	if err != nil {
 		w.Close()
+		releaseLock(lock)
 		return nil, fmt.Errorf("store: open read pool: %w", err)
 	}
 	r.SetMaxOpenConns(4)
 
-	s := &Store{w: w, r: r}
+	s := &Store{w: w, r: r, lock: lock}
 	ctx, cancel := context.WithTimeout(context.Background(), opTimeout)
 	defer cancel()
 	if err := s.migrate(ctx); err != nil {
@@ -81,12 +95,36 @@ func Open(path string) (*Store, error) {
 func (s *Store) Close() error {
 	err1 := s.w.Close()
 	err2 := s.r.Close()
+	releaseLock(s.lock)
 	if err1 != nil {
 		return err1
 	}
 	return err2
 }
 
+// acquireLock takes a non-blocking exclusive flock. Kernel releases it on
+// process death (kill -9 включительно), so a crashed run never wedges the
+// project.
+func acquireLock(path string) (*os.File, error) {
+	f, err := os.OpenFile(path, os.O_CREATE|os.O_RDWR, 0o644)
+	if err != nil {
+		return nil, fmt.Errorf("store: open lock file: %w", err)
+	}
+	if err := syscall.Flock(int(f.Fd()), syscall.LOCK_EX|syscall.LOCK_NB); err != nil {
+		f.Close()
+		return nil, fmt.Errorf("store: project database is in use by another tmctl process (lock %s): %w", path, err)
+	}
+	return f, nil
+}
+
+func releaseLock(f *os.File) {
+	if f == nil {
+		return
+	}
+	_ = syscall.Flock(int(f.Fd()), syscall.LOCK_UN)
+	_ = f.Close()
+}
+
 func opContext() (context.Context, context.CancelFunc) {
 	return context.WithTimeout(context.Background(), opTimeout)
 }
diff --git a/backend/internal/store/store_test.go b/backend/internal/store/store_test.go
index f44d216..ca530b6 100644
--- a/backend/internal/store/store_test.go
+++ b/backend/internal/store/store_test.go
@@ -188,3 +188,23 @@ func TestMigrateIsIdempotent(t *testing.T) {
 		s.Close()
 	}
 }
+
+func TestSingleProcessOwnership(t *testing.T) {
+	// Второй Open того же проекта обязан отказать: параллельный процесс
+	// (report во время translate) recovery-проходом обнулил бы живые резервы.
+	path := filepath.Join(t.TempDir(), "test.db")
+	s1, err := Open(path)
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer s1.Close()
+	if _, err := Open(path); err == nil {
+		t.Fatal("second Open must be denied while the project is locked")
+	}
+	s1.Close()
+	s2, err := Open(path)
+	if err != nil {
+		t.Fatalf("after Close the lock must be free: %v", err)
+	}
+	s2.Close()
+}