agent: move the loop prompt's memory into the corpus, and let it push

The autonomous agent's loop prompt had grown to ~9 000 words of accumulated
findings, refutations and traps. That is a symptom, not a style: it was doing
the job the corpus should do, in the one place that dies with the container.
Three concrete failures followed from it, and each gets a structural fix rather
than a louder instruction.

1. IT REPEATED WORK IT HAD ALREADY DONE.
   The "do not revive" list and the method traps existed ONLY in the prompt, so
   nothing pointed a fresh iteration at them. Extracted verbatim into two
   tracked files:

     docs/re/REFUTED.md  105 claims tested and dead, grouped by subject so a
                         grep for your noun finds the neighbourhood
     docs/re/METHOD.md   the traps already paid for -- controls, inference,
                         searching, reading data, runtime

   Both are linked from INDEX.md, and the loop prompt now opens by requiring
   them to be read. This is the fix for "re-derived something already known":
   the knowledge is now where the next iteration looks, not in a context window.

2. IT FORGOT TO ARM THE NEXT WAKEUP.
   The prompt tried to solve this by shouting at itself in the first line. The
   real fix is to stop asking: `loose` now defaults to a FIXED interval (45m),
   so the harness owns the cadence and a forgotten ScheduleWakeup cannot end the
   run. SYLPH_LOOP_INTERVAL= (empty) restores self-pacing.

3. IT COULD NOT PUBLISH, SO THE WORK ONLY EXISTED IN THE CONTAINER.
   New `push-work`, plus a read-only credentials mount
   (SYLPH_GIT_CREDENTIALS, default ~/.sylph-git-credentials). It pushes the
   CURRENT branch only, refuses anything that is not auto/*, and never
   force-pushes -- so the consolidated line stays a human's decision and a
   confused iteration cannot rewrite history. The loop prompt now requires a
   push on every iteration that commits, rather than at the end of some longer
   arc, which is exactly when a container dies.

The prompt itself drops from ~9 000 words to 85 lines and, more importantly,
stops accumulating: findings go in the corpus, and the prompt points at it.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

docs/re/INDEX.md

@@ -2,6 +2,11 @@
 
 Confidence: ✅ `CONFIRMED` · 🟡 `PROBABLE` · ❔ `HYPOTHESIS`. See [README](README.md).
 
+Also durable, and worth reading before proposing anything:
+[`REFUTED.md`](REFUTED.md) — what has already been tested and died ·
+[`METHOD.md`](METHOD.md) — the traps this corpus has already paid for ·
+[`BACKLOG.md`](BACKLOG.md) — what is still open.
+
 Formats we've already reversed are, for now, **documented by their parser + disc round-trip
 tests** (the executable spec) rather than a prose file — the "Spec" column points there.
 Promote to a prose `structures/…md` file when a format needs behavioural notes beyond layout.

docs/re/METHOD.md

@@ -0,0 +1,105 @@
+# Method traps already paid for
+
+Each line cost an iteration at least once. They are general — they are not about
+Sylpheed, they are about how this kind of measurement goes wrong.
+
+Like [`REFUTED.md`](REFUTED.md), this list had been living in the autonomous
+agent's loop prompt, i.e. nowhere durable. See [`README.md`](README.md) for the
+✅/🟡/❔ confidence convention itself.
+
+## Controls
+
+* **Every result needs a control. A control that fails kills the instrument.**
+* **Run the known-positive through a new filter FIRST.** Three filters have been
+  killed by their own control. When one fails, **read the known-good's
+  disassembly** before assuming a shape.
+* **A measured negative is a result** — but a negative is only as strong as the
+  route you ran, so **state its reach**.
+* **A null result needs its cause shown to have happened.**
+* **A result with NO unknowns is suspicious.**
+* **Census the whole set; always run the other population as the control.**
+  **Zero partials is stronger than a majority.**
+* **A 2×2 partition is the sharpest general tool** — both off-diagonals empty is
+  a law.
+* **Re-derive a doc's own numbers as the control.**
+
+## Inference
+
+* **Never conclude from ONE sample.**
+* **A law proved on one population is a hypothesis on the next.**
+* **Finding one exception does not imply a family.**
+* **Consistency is not proof. A suggestive coincidence is a coincidence until
+  measured.** **An analogy is not a measurement.**
+* **Same layout ≠ same instance.** **Same record-name set ≠ same object.**
+* **A marker is only proven by what it leaves out.**
+* **A high-confidence SCORE is not a high-confidence MECHANISM.**
+* **Knowing HOW MANY is not knowing WHICH.**
+* **Round numbers matching is weak evidence — unless you read the constant.**
+* **My own last-turn result is a hypothesis too.**
+* **A global partition can understate a per-owner one.**
+* **A residual is measured against a population — name it.**
+
+## Searching and tooling
+
+* **A search that returns thousands has no power; state the reach.**
+* **A substring match is not a hit.** **A regex miss looks like a null result —
+  print one raw sample before believing a zero.**
+* **A derived table can be a cross product — measure its shape first.**
+* **After refuting an instrument, sweep everything that depended on it.**
+* **Before measuring how wrong a tool is, read what the tool actually does.**
+* **The instrument must pass its own control.**
+* **Classify a bulk before mining it. The residual is the prize.**
+* **Rank by similarity — the cliff is the finding.** But **read the values
+  before trusting the rank.**
+* **Grep the nouns before designing the experiment — and believe it.**
+* **Grep gives you a file list — read *every* file on it.**
+* **The answer is often already in the doc that owns the subject — read it end
+  to end.** A 🟡 often names its own route.
+* **Before re-trying a blocked idea, check whether the blocker's own doc already
+  tried it.**
+* **Ship a regenerator with every artefact.** An artefact that moves by a pure
+  reorder is a tool bug.
+* **Never print per-entry lines from a disc-wide sweep — aggregate.**
+
+## Reading the data
+
+* **Read what a loader NAMES, not where it stores.**
+* **A field the disc never values still gets named by the loader.**
+* **An indexed read beats a deduped-pool adjacency read.**
+* **Check the whole string set, not the one matching word.**
+* **A dict keyed by record name across a multi-entry pak is a lie.**
+* **A set-difference over names hides reuse — join per USER.**
+* **A self-index names records, not files.**
+* **Case-insensitive hashing means two spellings can be one entry.**
+* **An "unresolved" name may be the wrong kind, namespace or prefix — or part of
+  a cut asset.**
+* **A garbled value may be a real string in another encoding.**
+* **Two of my own counts disagreeing is a grammar clue.**
+* **A game's own typo is a join key.**
+* **A bias constant in the code is a join key.**
+* **A prefix trap: enumerate maximal `[A-Za-z0-9_]` runs, not `startswith`.**
+* **Re-deriving a format is not a finding — asking whether its values *resolve*
+  is.**
+
+## Mechanics that have bitten
+
+* **Never hand-convert a decimal VA — print `hex()`.**
+* **`grep -c` counts LINES** — use `grep -o | wc -l`.
+* **`Counter.most_common()` tie-breaks by insertion order — use `sorted()`.**
+* **Raw grep cannot see inside compressed pak entries.**
+* **Commit messages go in a file** (`git commit -F`); a literal `|` in a table
+  cell needs escaping; `git log --all -- <path>` can hang.
+
+## Runtime / emulator
+
+* **Look at the PNG** — and check its dimensions.
+* **"Animating" is not "still in a mission".**
+* **Dedup entity enumerations by position value.**
+* **Do not diagnose timing or liveness under gdb.** `ps %cpu` is cumulative.
+* **Classify screens by whole-image statistics, not named pixels** — a named
+  pixel is only valid while the image sits at a known place, and nothing errors
+  when it moves.
+* **Do not poll faster than the guest updates** — it manufactures a clean curve
+  out of noise.
+* **A probe that never performs the action will "prove" the action does not
+  exist.**

docs/re/REFUTED.md

@@ -0,0 +1,185 @@
+# Refuted — claims that were tested and died
+
+**Read this before proposing a hypothesis.** Every line below was believed at
+some point, measured, and found false. Reviving one costs a whole iteration and
+produces nothing.
+
+This file exists because the list had been living in the autonomous agent's
+*loop prompt* — the only copy, lost the moment the container was. A refuted
+result is a real result; it belongs in the corpus like any other.
+
+**Format:** each entry is the claim as it was believed. Where the true answer is
+known it follows after `→`. Grouped by subject so a grep for your noun finds the
+neighbourhood, not just the line.
+
+---
+
+## Offsets, structs and the progress singleton
+
+* `position = instance − 0x12c` → refuted.
+* `+0x29d0` → refuted.
+* "an offset intersection can find a struct's consumer" → **only for LARGE or
+  unusual offsets.** Small ones have no power (`+184`: 301/351/115 hits).
+* "a `+1956` store means a progress write" → writes go through the COPY, not
+  direct stores. 9 direct stores, none of them a progress write.
+* "the singleton-global filter can find progress writers" → it fails its own
+  control.
+* "the progress copy destination is an `r1`-relative stack local" → it is a
+  **frame register**. The `r1` assumption returned 0 for all 21 candidates
+  *including the known-good* — the filter was killed by its own control.
+* "word B's writer also stores the Time/Points record" → it does not.
+* "the debriefing records the metric with the clear bit" → 44 calls, exactly two
+  strings (`DEBRIEFING`, `BASE_INFO`), no `Time`, no `Points`.
+* "`0x820AF030` holds live state" → all 384 words constant; it is a
+  spawned-entity record, not live state.
+
+## Screens, classes and RTTI
+
+* "the RTTI route can name the anonymous classes" → 1 150 vtables: 1 150
+  `ANON_`, 0 `rtti_present`, 0 base classes.
+* "the sibling vtable methods name the class" → they cannot.
+* "`xrefs` can name the callers of a vtable method" → no.
+* "the `ind_call` refutation voids existing corpus claims" → damage bounded,
+  4/4 caller claims verify. But **`xrefs.ind_call` is a CROSS PRODUCT** — always
+  filter `kind='call'`.
+* "`BASE_INFO` marks the 5-slot screen family" → it discriminates
+  screen-config from table-read, 9/9 vs 10/10.
+* "a high key count means a rich screen" → `sub_82297550` / `sub_822A2F00`'s 27
+  "keys" are coordinate pairs, i.e. a layout table.
+* "`EX_` = the CHALLENGE-mission debriefing" → `EX_` is **EXTRA**, mission-kind
+  3.
+* "the `EX_` selection has not been shown" → it has: `[[obj+4]+184] == 3`.
+
+## Stages, missions and the challenge set
+
+* "the disc's stages are numbered 1..28" → S01–S16 story, S17 **cut**, S18–S23
+  tutorials, S24–S29 challenge.
+* "S24–S29 are story missions" → they are the challenge missions.
+* "the challenge missions have their own maps" → they reuse
+  `GP_MAIN_GAME_E.pak`'s stage records.
+* "the challenge `REQUIREMENT` values are 16,25,26,27,29" → the chain is
+  16→24→25→26→27→28.
+* "the `Extra0n` family shares one leaderboard metric" → `RECORD_TYPE` is
+  per-stage: 3 Time / 3 Points.
+* "stage = filled SHAB count + 1" → refuted.
+* "the first TRIGGER is always the point of no return" → refuted.
+* "`EnumUnit_S14.tbl` might be missing" / "S14's 13 are a manifest omission" /
+  "asteroids are exempt from the manifest" → S14's 13 are **dangling
+  deployments**. NEEDS-HUMAN: fly S14.
+* "`StageMessageSet_S02.tbl` is not in the pak" → it is.
+* "`S28_p1` has an asteroid volume with no definition" → refuted.
+* "`test_s8p1_asteroid.tbl` is test-only" → refuted.
+* "the settings family has 28 or 29 objects" → 24.
+
+## ISL / mission scripting
+
+* "the bytecode is in the `.embsec_` sections" → refuted.
+* "only 31 built-ins take a unit" → refuted.
+* "`sub_8230C398` is the message pump" → refuted.
+* "`bus+8216` is the subscriber registry" → refuted.
+* "the ScriptPhase vtable is ≥200 slots" → 113.
+
+## IDXD, paks and naming
+
+* "IDXD record keys are `name_hash`" → record keys are **`tag_hash`**
+  (case-SENSITIVE); `name_hash` is case-INSENSITIVE and used for pak keys.
+* "pak TOC order is stage order" / "TOC order is semantic order" → it is not.
+* "the executable holds the asset names" → the image names **no data value at
+  all**; that route is powerless.
+* "the image might name a data VALUE" → powerless.
+* "the XPR2 manifest names hash to the DefTables tables" → refuted.
+* "the `DefTables` model names are unreachable" → reachable via the `Enumerate`
+  declaration tables (1 413/1 425, 99.2 %).
+* "the `GP_MAIN_GAME_*` unnamed block is undiscovered data" → refuted.
+* "each `GP_MAIN_GAME_*` `Enumerate` object declares something" → refuted.
+* "`GP_HANGAR_ARSENAL` is missing data tables" → refuted.
+* "the `Enumeration` self-index can name objects" → a self-index names
+  **records, not files**.
+* "a per-pak prefix might close the 2D blocker" → no.
+* "the `+` paths might name the 2D or `GP_READY_ROOM` keys" → the `+`-dictionary
+  route is exhausted, 0 of 1 817.
+* "the `game:\` paths are unresolved" → refuted.
+* "a set-difference over file names can see reuse" → it cannot; **join per
+  USER**. Per-pak copies are ×6.
+
+## Units, weapons, effects and assets
+
+* "`Generic` (394) is the unit datasheet" → refuted.
+* "a loadout's `Arm1` names an item" → it names a **hardpoint slot**
+  (`Turret_NNN`), 59/59.
+* "`EnumUnit` and the unit datasheet share a vocabulary" → they do not.
+* "the roster is the `Generic.Model` set" → roster 40, `Generic.Model` 46,
+  `GameResourceID` 480 — three vocabularies.
+* "every unit ID is `UN_<l>###_<FACTION>_<name>`" → the grammar is
+  `UN_<letter>###_[<subkind>_]<FACTION>_<name>`.
+* "`_EXn` is the `Extra0n` index" → three different `EX` vocabularies exist.
+* "the only two `_EX5` names on the disc are the AA gun and the DeltaSaber" →
+  refuted.
+* "running the tutorial will instantiate the `_Ttrl` weapons" → refuted.
+* "the disc has exactly three `EnumWeapon` tables" → four.
+* "the `wep_NN` package gaps are unshipped weapons" / "`wep_85` is the tip of a
+  family" → `wep_85` is the **only** declared-but-unshipped asset (59/0/1/26).
+* "nothing is deployed without being declared" → refuted.
+* "effects are one namespace" → refuted.
+* "the 58 undeclared effect names are missing assets" → refuted.
+* "all five orphan effects are unshipped" → refuted.
+* "`eff_f0002` ships in `Base.xpr`" → refuted.
+* "`Base.xpr` holds more bound effects than `ptc_pack`" → refuted.
+* "the 34 unlocated are a scatter" → refuted.
+* "the 9 unlocated might be under another prefix" → refuted.
+* "`ptc_pack` has 532 names" → 727.
+* "the `_e`/`_f` law is effect-FIELD-specific" → it is the **faction law**,
+  564/564.
+* "a disc-wide `.xpr` byte search can show an effect is ABSENT" → it cannot.
+* "`rot_n001` is on the disc" → refuted.
+* "`rou_f004`'s mesh is in `Stage_S28.xpr`" → it is in `DeltaSaber_A.xpr`.
+* "`parent` + `_all` + `_child` is the composite-model convention" → refuted.
+  `_hangar` **is** real (59 of 166, 59/59 with a bare twin); `_all`/`_child` is
+  not.
+* "`Motion_guard_start` has no damage variants" → refuted.
+* "`CoverArea` bits 2 and 3 are mutually exclusive" → refuted.
+* "the 27 unresolved `NamePlate` values are missing objects" → refuted.
+
+## LOD, background and misc tables
+
+* "`EnumLODSet_*` is a per-stage family" → `EnumLODSet_test.tbl` serves 17
+  stages; 17+5+1 = 23.
+* "there are 8 orphan LOD tables" → 6.
+* "the orphans are stale copies of `_test`" → refuted.
+* "S25 is absent from the `DefTables` LOD families" → refuted.
+* "`BackGroundID` has no referent anywhere" → it is an **identity**.
+* "`BackGroundPackage == BG_<id>.xpr`" → refuted.
+* "`<X>ID` + `<X>Package` is a convention" → refuted.
+* "`Placement_*` / `RouteTest_*` are unattached" → refuted.
+* "the `AsteroidDefinition` join does not reproduce by hash" → it does.
+* "the 8-value frame is a new finding" → it was already in the corpus.
+
+## Loaders, config and tuning
+
+* "the config reader is XML" → INI.
+* "`sub_822F9498` is the unit-definition loader" → it is `PlayerParams`'s.
+* "`sub_822AE628` reads the main-game `Tweak`" → refuted.
+* "`sub_8230D1F8` is a rank table" → it is the stage-settings loader.
+* "`sub_82286BC8`'s key list is new" → refuted.
+* "`sub_825F2CF0` / `sub_825F2F88` read a post-processing table" → refuted.
+* "`Booster` is a new schema" / "`Booster` is the player craft's flight
+  envelope" → refuted; nothing selects `Booster`.
+* "the `AnalogRevice`/`Tweak` block is unreachable" → reachable
+  (`sub_821A6CF0`, base `0x820A1630`).
+* "a 0-xref string block has no reader" → refuted.
+* "the AI table was NEEDS-HUMAN" → refuted.
+* "the `PG*` HUD names are undocumented" → they are documented.
+* "a base-solver row identifies a FUNCTION" → it does not.
+* "a 64K-boundary base is low confidence" → **inverted**; it is high
+  confidence.
+* "the 0x820B0000 cluster is a false positive" → refuted.
+* "a pointer to a function in the image implies a registry" → refuted.
+  `.pdata` is not a registry.
+* "the 13 player-facing chatter tables are the WINGMAN tables" → refuted.
+* "the 8 undeclared chatter tables are tutorial chatter" → refuted.
+* "other datasheets ship a schema too" → refuted.
+
+## Encoding and text
+
+* "every IDXD string value is ASCII" → 6 non-ASCII values of 99 328.
+* "`文字列` is a dev placeholder" → they are Shift-JIS **type words**.