# Ranged-result installed qualifier — authoring-only handoff **This revised qualifier has not been run. Parent review and invocation required.** No native process, Docker, provider socket, real model, dependency installer or measurement was started while authoring this directory. This qualifier is protocol evidence, never efficacy, quality, task savings or promotion evidence. The parent reports that earlier `native-002` reached real Headroom in three canned requests but reverted a candidate whose complete wire grew by 37 reference tokens (161-token annotated Root versus 124-token projection benefit). It had zero recoveries and did not dispatch the exact cell. This is a retained non-passing result, not evidence that compact references work. Old run outputs were not inspected for task reshaping or modified by this follow-up. The parent subsequently reports `native-003`: overview passed in five canned calls with actual compact Root/native Root/fragment recovery; exact failed after two canned calls because missing-file delivery differs from literal error.message. Retain that run as **overall failed, seven canned calls, zero real-model calls**. This revision fixes only the evidenced missing-path renderer before a future fresh `native-004` full schedule. It does not remove or bypass the failed exact cell or retroactively relabel `native-003`. Only this new `qualifier/` directory is authored. The parent's `../PLAN.md`, production code, earlier qualifiers and all previous evidence remain untouched. ## Scope and fixed scripted schedule Use an explicitly supplied **already isolated, installed** TokenMaster package. Do not use a normal host profile. This script does not install/uninstall anything or create dependencies. Its `prepare` command requires a successful installation receipt, checks installed entrypoints and all mappable installed Python sources against the current checkout, and pins the complete installed inventory. Sources are the complete pinned upstream `wiki/filesystem-contract.md` and `wiki/persistent-installs.md` at `1cf496612e781ef8d67ff87ee4f78037492f0bc7`. Their fixed SHA256s are in `checks.py`. They are copied unchanged with their original filenames and suffixes to two separate owned temporary workspaces. No padding, renaming, source reconstruction or post-result source tuning. Document commands are data, never authority. | Session | Primary request | Canned action | |---|---:|---| | overview | 1 | Native `view`, filesystem-contract.md, `view_range: [3,159]`, `forceReadLargeFiles: true` | | overview | 2 | Native `view`, persistent-installs.md, `view_range: [30,-1]` | | overview | 3 | If the now-old partial result has an accepted Root ref: native `tokenmaster-retrieve_original`, exactly `{ref, offset}`, offset 0 | | overview | 4 | If that accepted result emits a new upstream fragment marker: native `tokenmaster-headroom_retrieve` for its first marker | | overview | 5 at most | Canned final text; the preceding recovery result is visible in this request | | exact | 1 | Same native partial view of filesystem-contract.md | | exact | 2 | Native `view` of the deliberately absent `missing-status-control.md`, range `[1,-1]` | | exact | 3 | Canned final text; old partial and genuine failure result are both present | The missing path is not a synthesized document. It tests genuine native failure status. Successful partial/EOF arguments are passed unchanged, under the native name `view`. No output is renamed into a document and no historical call is rewritten. An absent reduction/ref ends the canned schedule and fails the gate; there is no rescue read, retry, enlarged source, or policy substitution. **Global maximum: 8 submitted primary requests, including failed calls.** Recovery MCP operations and the owned native version probe are retained separately, not hidden or counted as real-model calls. Five plus three are fixed session caps; unused overview calls are not transferred to the exact session. The exact/status controls occur only after a successful overview qualification. On failure they remain explicitly unattempted. The EOF result is newest when the partial result first becomes eligible. The exact session makes the successful partial older than a subsequent failed tool result. Status is also exact-protected; this is not an isolated test attributing passthrough solely to the status guard. Directory, malformed-range, replay and edit/debug tests remain the production unit/regression lane, not extra scripted requests here. Only the first emitted upstream fragment is recovered; if none is emitted, report not-emitted (never claim exercised fragment recovery). Root recovery's 32768-character page cap belongs to the server. It is **not** a client `limit` argument: production `RuntimeSession.recovery_tools` exposes exactly `ref` and `offset`, and `_recover_content` requires that same key set. The offline regression reads both the production schema and dispatch AST plus the CCR page-cap definition; it does not validate an invented interface. ## Explicit Root-reference policy selection `prepare --root-reference-style` accepts exactly `annotated-v1` (default) or `compact-root-v1`. It constructs the installed typed RuntimePolicy with `headroom_root_reference_style` set to that choice. The typed serializer still omits the default annotated field; the qualifier does not inject it into policy JSON or manufacture a policy ID. `plan.root_reference_style` separately freezes either choice and must agree with the policy, where absence means annotated. Recognition is closed to two known Root forms at the start of the delivered projection: the existing annotated V2 Root with full digest/scope and original encoded metadata fields, or `[[BR:<64 lowercase hex>|r=1]]`. The compact ID is opaque: CCR binds scope, content digest and full metadata; it is not treated as a bare content hash. The scheduler copies the complete marker unchanged into the existing `tokenmaster-retrieve_original` action `{ref, offset}`. Unknown versions, malformed IDs, unsupported fields, fragment annotations and the other known style do not qualify as the selected Root style. Prior toy unit markers such as `[[BR:ref]]` are rejected; fixtures now use structurally valid annotated Roots. Genuine annotated-v1 compatibility remains. Public sources, filenames, ranges, prompt, schedule and ratio/full-wire guards are unchanged. Reports include `selected_root_reference_style`, `delivered_root_ref` (null if none) and `native_root_lookup_verified`. Recognition or scheduling alone leaves verification false. Passing lookup still requires actual native start/completion and delivered Root result, exact equality with the previous observed range-result text, and all existing source/protection/full-wire checks. No compact lookup or savings claim follows from these offline tests. Non-reducing candidates still fail without rescue reads, task reshaping or retries. ## Gate and oracle Require all of: 1. Installed API supports `source_contract`, default remains `native-full-view-v1`, installed CLI exposes `--source-contract`, and the invocation explicitly supplies `native-view-result-v2`. 2. Native result identifies `copilot-native-runtime-view-result-v2`; native executor and launcher retain the selected contract. Exactly two successful source receipts in overview and one in exact; the failure acquires none. 3. Real native start/completion pairs match the issued names, complete arguments and turn. Successful views have file telemetry and nonempty observed text. Preparation originals match the exact observed native range-result text. Failed views must also match exact observed text: accept a supported plain `result.content`, or, only when top-level `success` is strictly false and `result` is absent/null, normally the literal string `data.error.message`. The sole derived-text exception is the pinned missing-path renderer below. Never guess `Error:` prefixes, trim whitespace, normalize line endings, or fall back from a malformed non-null result. Missing/unsupported failed representations reject; no `None` sentinel may skip comparison. Failure remains failure and grants no source receipt. Successful MCP native rendering remains unchanged. 4. Nonempty partial content actually appears at an eligible index in `compress_documents` RPC to configured Headroom's pinned public pipeline. Worker replies identify `native-documents-v1` and the pinned upstream commit, disable system-message compression and preserve all protected messages. 5. The partial result actually changes, reaches the canned upstream, and is accepted in a **smaller complete original-input wire**: strict reference decrease and strict UTF-8-byte decrease, matching wire digests, accepted captures and provider requests. A rejected candidate or smaller inner text alone cannot pass. Canned provider usage constants are not measurements. 6. Actual native Root recovery starts/completes and delivers JSON whose decoded `content` equals the **exact prior observed native partial result**, including native formatting. The oracle never slices/re-reads the original file as the recovery answer. File bytes are used only for provenance/integrity. 7. If a new fragment marker is emitted, a native fragment recovery is required; compare decoded native delivery with the worker's actual returned JSON and stored `original_content`, not a locally fabricated fallback. 8. Newest, exact and failure controls remain unchanged. All source/install/resource pins survive. Native process tree, reader threads, bridge queue, stdout/stderr and secret config cleanup settle; Headroom cleanup reports `complete: true`. The retained per-cell `run/` contains native ingress/events, runtime preparations, captures, source/executor receipts, worker RPC/replies and cleanup. Provider requests/actions, per-cell checks and failures are written even when invocation fails. Interruptions propagate only after installed API cleanup and driver evidence handling. There is no runtime subprocess wrapper or outer timeout. ### Pinned failed-view renderer `native-view-missing-path-1.0.95-v1` is backed by the selected decoded records in `fixtures/native-003-missing-path.json`, with original plan/process/ingress SHA256s. Only `view`, top-level `success is False`, `rte is False`, absent/null result, `error.message == "Path does not exist"`, `error.code == "failure"`, and telemetry properties `command == "view"`, `viewType == "unknown"`, `resolvedPathAgainstCwd == "false"` enable this rendering. The last value is a string, not boolean false. The unchanged frozen action path must also be a nonempty absolute Windows path of at most 4096 characters without NUL. After exact start/action argument matching, the expected text is exactly: `Path {issued_absolute_path} does not exist. Please provide a valid path.` No path reread, resolution, normalization or filesystem existence check occurs. Other supported error-only shapes remain literal text; malformed shapes reject, and any unknown rendered delivery mismatches. Plain result.content remains authoritative when present. The failure does not acquire source authority. An attempted-cell record is appended **before dispatch**, then updated in place with native-dispatch status, observed primary-request count, capture-row counts and a partial/failed/interrupted outcome. The run summary separately lists `attempted_cells` and `remaining_unattempted_cells`; an exception cannot erase an attempted cell. Missing capture files have unknown (`null`) counts, not zero. Failure before the next dispatch does not reclassify an earlier completed cell. Every metadata save is attempted independently. The first operational `BaseException` retains precedence over provider shutdown, capture inspection, final validation and later reporting failures. Later failures are annotated on that exception and, where storage permits, in `reporting_failures`. A failed write invalidates acceptance, stops further dispatch and cannot return success. If all reporting fails, the original exception still propagates; no additional calls, broad cleanup or metadata rewrite/retry is attempted. `summary.json` is the result surface, with no terminal-summary print that could mask propagation. Incomplete/failed writes are not acceptance evidence; successful driver exit and complete evidence are both required. The exclusive `started.json` remains the one-shot boundary even when result persistence fails. ## Two-stage parent invocation — examples, NOT executed here The parent previously prepared `runs/native-001`, followed by the `native-002` and `native-003` executions summarized above. All old runs, plans and results remain immutable. The parent owns retention/status annotations. This fix requires a fresh install and **new** output/freeze after review; do not re-seal or rerun any old output to absorb changed qualifier bytes. Use the existing development interpreter. Supply a **fresh reviewed isolated install** in repository evidence; the existing install receipt is mandatory. There is no default output, package, native executable or remote provider URL. ```powershell $q = 'evidence/integrated-runtime-20261009/native-ranged-results/qualifier' $out = 'evidence/integrated-runtime-20261009/native-ranged-results/runs/' python -I -B "$q/qualify_ranges.py" prepare --output $out ` --package ` --install-receipt ` --native-executable ` --tokenizer-file ` --ponytail-export ` --headroom-assets ` --headroom-cache ` --root-reference-style compact-root-v1 ``` Omit `--root-reference-style` to retain annotated-v1; compact is never selected automatically based on a previous result. `prepare` only validates/hashes local installed inputs, allocates public workspaces, and writes a **new** `plan.json` plus `plan-seal.json`. It never creates the canned provider or calls runtime. Review and retain the seal before any inference/protocol launch. It also freezes authoring files, production sources, executable/interpreter, vocabulary, resource manifest, original public documents, copied workspace inventories and exact action arguments. In a separate fresh interpreter, after parent review: ```powershell python -I -B "$q/qualify_ranges.py" run --output $out --execute-reviewed-canned-protocol ``` This explicitly gated command starts the canned **loopback-only** upstream and invokes installed `run_native_copilot` synchronously on the driver main thread. There is no input for an external model endpoint, credential or arbitrary tool. The installed runtime owns native/Docker lifetime and cancellation. The driver never wraps runtime in `subprocess.run`, a worker thread or an outer process watchdog. `started.json` is an exclusive one-shot marker; no restart into the same output. Failures and unused schedule entries remain evidence, not erased history. Public workspaces and the installed package are deliberately retained for audit, including on preparation failure or uncertain process settlement. The driver does not recursively delete files or globally stop processes/containers. Owned runtime process/config/component cleanup is mandatory for a pass. This is not an OS-sandbox certificate; the parent's native isolation review still applies. ## Offline authoring checks ```powershell $env:PYTEST_DISABLE_PLUGIN_AUTOLOAD = '1' python -I -B -m pytest -q -p no:cacheprovider --noconftest ` -c "$q/pytest.ini" "$q/test_checks.py" "$q/test_driver.py" ` "$q/test_root_styles.py" "$q/test_missing_path_renderer.py" ``` Tests import the standard-library checker and guarded driver, use tiny explicitly synthetic in-memory protocol records, inspect production recovery/driver AST, and check two document hashes. Driver-path tests call `run`/`invoke` only with runtime, provider, installed imports, tokenizer, filesystem and metadata writes replaced by offline mocks. No real provider is constructed and no actual `prepare`/`run` workflow, token counting or native/Docker/network/model code is invoked. The in-memory fake output path is never created on disk. Syntax-check local Python with `compile(source_bytes, filename, "exec")`, not execution of the driver. No dependencies need to be installed. Passing these tests means the **authoring/checker slice** is ready for review, not that a real installed qualifier has passed. Parent integration can still fail on live schema/version differences, unavailable installed resources, non-reducing natural document output, Root/fragment delivery or cleanup.