dtf-system/docs/ROADMAP.md

DTF System — Working Roadmap

Internal engineering tracker. Not a client document, not a promise sheet. The client-facing narrative lives in docs/roadmap-cliente.pdf and in the weekly report, which is a deliverable and is not versioned here.

Update the Current step line and the item status every time something moves. Add new findings at the bottom of the relevant block rather than rewriting history.

Current step: Block 0 closed; Block 2 closed except 2.9–2.11; 4.1, 4.2, 4.5, 5.1, 5.3, 5.4 and 5.8 done. Remaining work needs decisions (3.1, 3.2, 3.3) or client inputs (1.1, 1.2, 2.10). Block 1 still waits on client inputs for 1.1/1.2.

Paths in closed items are written as they were when the finding was made. The repository was laid out by role on 2026-09-21 (local/ became app/, with tests/, ops/, infra/ and web/ beside it); the history is left as recorded rather than rewritten.

Last audit: 2026-09-18, full read of app/, dtf-site.html, deploy/, .gitea/, docs and legacy prototypes. Findings below carry their audit IDs.

Status	Meaning
[ ]	not started
[~]	in progress
[x]	done and verified
[?]	blocked on a decision (product or client), not on code

---

Block 0 · Broken right now

Nothing in this block is optional. Until it is closed, the system cannot be demonstrated, and the week-1 claims cannot be defended.

[x] 0.1 — API returns 500 on every session, login and registration (F1)

local/auth.py:46 calls os.environ and the module never imports os. Reproduced: NameError: name 'os' is not defined. /api/session calls new_session() whenever there is no cookie, so the Site checkout bridge, cart recovery and the customer portal all fail on first visit. Introduced in e3e37f6.

• Add import os to local/auth.py.
• Move the COOKIE_SECURE read to a module constant so it is evaluated once.
• Accept: a fresh browser hits the Site and /api/session returns 200 with a cart_scope; local/smoke_test.py and local/workflow_test.py pass.

[x] 0.2 — "Arquivo por metro" is unreachable in practice

Two independent causes, both from 2026-09-18 commits. Verified in a browser.

a. Every entry point hard-routes to loose artwork (483a083)

Control	Currently opens
Nav "Impressão DTF"	avulsa
"DTF Têxtil · 57 cm"	avulsa
"DTF UV · 28,5 cm"	uv
Hero "Enviar minha arte"	avulsa

Only the price card reaches file. Revert the three data-modo-cta attributes so navigation links land on the product chooser, not on a product.

b. Dropping a PNG/JPG in file mode silently switches the order (F25)

dtf-site.html sel() — from abrir('file'), dropping imagem-teste.jpg gives modo: "avulsa", header "Artes avulsas", price R$ 29,90/m. Meanwhile the same screen says "Arraste suas folhas montadas · PNG, JPG ou PDF", marks "PNG, JPG ou PDF · a partir de R$ 14,90" as the recommended path, and sets input.accept=".png,.jpg,.jpeg,.pdf". The page invites the drop and then reprices the order 50% higher.

The escape hatch #imagemComoFolha sits above the drop zone as small text inside an informational notice, defaults to unchecked, and must be ticked before the drop. After the switch fires, pintaModo() sets trocarParaAvulsa.hidden = true, so the checkbox disappears and there is no way back in place.

Net effect: the only formats that survive file mode are PDF/TIFF/PSD/AI/CDR — the path the UI itself marks as the worse option. A mixed drop (JPG + PDF) is rejected and accepts nothing.

• Make the ready-sheet choice an explicit two-option control inside the drop area, styled like the existing .cam selector — not a checkbox in a notice.
• Stop advertising PNG/JPG in the by-metre drop zone while rejecting them.
• When a switch does happen, show the price change and offer one-click undo.
• Accept: a customer can complete a by-metre order with a PNG from any entry point, and no product/price change ever happens without a visible confirmation.

[x] 0.3 — Ready-sheet declaration is unverified and worth money (F22 related)

Ticking #imagemComoFolha is an honour-system claim that moves the price from R$ 29,90/m to R$ 19,90/m (R$ 14,90 with a good grade). medirFolha then derives sheet height purely from aspect ratio × 57 cm — the original bug, now opt-in.

Measured with imagem-teste.jpg (466 × 659 px):

Route	System behaviour	Billed
Ticked	treated as a 57 × 80,6 cm mounted sheet	1 m × R$ 19,90 = R$ 19,90
Not ticked	20 cm wide, packs to 28,3 cm of film	1 m × R$ 29,90 = R$ 29,90

• Validate the claim: declared width must be ≈ film width (57 / 28,5 cm) at a plausible DPI before the sheet model is accepted.
• Re-check server-side in /api/operator/quotes/{id}/approve before pricing.
• Accept: a small single artwork declared as a ready sheet is rejected with a clear message; a genuine 57 cm sheet passes; the operator sees the verdict.

[x] 0.4 — Documented local startup fails (F2)

LOCAL_SETUP.md says docker compose up --build with no .env. docker compose --env-file .env.example config exits 1: R2_ENDPOINT, R2_ACCESS_KEY_ID, SITE_DOMAIN, KANBAN_DOMAIN missing. docker-compose.yml became a production/R2 stack in e3e37f6; .env.example is still the MinIO one and there is no MinIO service left.

• Decide: keep one production compose and add compose.local.yaml with MinIO, or restore a local default. Recommend the former.
• Accept: a clean clone reaches a working Site + Kanban with the documented command, and LOCAL_SETUP.md matches what actually runs.

[x] 0.5 — App DB role shares the admin password (F3)

docker-compose.yml:65 sets APP_DB_PASSWORD: ${POSTGRES_PASSWORD} — the same value as dtf_admin. bootstrap.py grants the app role DML-only and then hands it a credential that also logs in as the owner. Anyone reading the API container env has admin on the database.

• Accept: distinct secrets; connecting as dtf_app with the admin password fails.

[x] 0.6 — A paid order showed the customer nothing (found while closing Block 0)

#checkoutStatus and #checkoutActions lived inside #carr, and the success path in checkout.js clears the cart (pedido=[]; limpaPaineis()) before writing the confirmation — .carr{display:none} then hid the panel holding it. Present since the first commit; it only surfaced once 0.1 made a payment reachable at all. Both elements now sit in their own always-visible .checkout container.

How Block 0 was verified

A live stack (compose.local.yaml), then the full suite:

Check	Result
/api/session on a cold browser	200 with cart_scope + session cookie
smoke · workflow · security · scanning	pass
retention · runtime security (in-container)	pass
artwork_browser_test.mjs	pass, updated to the new declared-product behaviour
browser_test.mjs end-to-end	pass — upload → quote → operator approval → paid order → all Kanban states
4 CI unit tests	pass

Ports 8090/8091/8010 were used; 8080 was held by an unrelated preview server.

---

Block 1 · Week 2 — committed to the client

From the report already sent. These are dated promises, not backlog.

• [~] 1.1 — Mercado Pago transparent checkout, signed and idempotent webhooks. The provider-independent half is built (2026-09-22): POST /api/payments/webhook verifies a signature before parsing, records every delivery under the provider's own event id, and applies it in one transaction. A duplicate is a no-op, a re-sent approval finds the order already there, and an approval whose amount disagrees with the reviewed quote is refused rather than shipped. Order creation moved to app/payments.py so the webhook and the local checkout cannot drift. Exercised end to end by tests/payment_test.py against a fake signer.

  What remains needs the client: a sandbox account, webhook administration, the event/status mapping and the refund policy. In code it is one adapter supplying create, verify and parse — nothing in the service changes.

• [ ] 1.2 — Real freight quotation. Blocked on client inputs (see PRODUCTION_INPUTS.md): source platform, credentials, origin CEP, services, packaging weight/dimensions per length, subsidy policy.

• [ ] 1.3 — Idempotent Tiny/Olist order creation with order-number traceability. Confirm endpoints, tag behaviour and rate limits first.

• [ ] 1.4 — Final print-file generation (see 3.2 — this is the same problem).

• [ ] 1.5 — Main Kanban production states consolidated.

• [ ] 1.6 — Block 0.2 + 0.3, promised as "início da próxima semana".

[!] The production compose currently blocks dev_paid (ENVIRONMENT != 'local') and ships only fake adapters, so the deployed system cannot take an order at all. 1.1 is what unblocks it.

---

Block 2 · Security — before any public exposure

[x] 2.1 — Rate limiting and audit logs are blind to the client (F5)

uvicorn runs without trusted proxy headers (forwarded_allow_ips defaults to 127.0.0.1; nginx is a different container IP), so request.client.host is nginx for every request. rate_limit('auth-source', ...) at 60/15min becomes a single global bucket — 60 failed logins lock out every customer — and every audit() record has no attacker IP.

• Set --proxy-headers with FORWARDED_ALLOW_IPS scoped to the nginx service, or read X-Forwarded-For explicitly at the edge.
• Accept: two clients on different IPs have independent buckets; audit rows carry the real IP.

[x] 2.2 — The public site throttles itself (F6)

local/app.py:115 — rate_limit('guest-sessions', ENVIRONMENT, 120, 900) is keyed on the environment name: 120 new visitors per 15 minutes site-wide (~8/min). Normal traffic 429s. Key per source IP (after 2.1) and raise the ceiling.

[x] 2.3 — deploy/stack.yaml cannot boot (F7)

It passes DATABASE_URL_FILE, AWS_ACCESS_KEY_ID_FILE, OPERATOR_PASSWORD_FILE, OPERATOR_USER. The code reads DATABASE_URL, AWS_ACCESS_KEY_ID, OPERATOR_PASSWORD, OPERATOR_EMAIL, and no _FILE loader exists. app.py:58 does os.environ['OPERATOR_PASSWORD'] → KeyError → 500 instead of 503.

• Implement local/secrets.py (the preflight already expects it) reading *_FILE with env fallback. Reconcile OPERATOR_USER vs OPERATOR_EMAIL.
• Accept: the stack renders and boots against Swarm secrets; missing operator config yields 503, not 500.

[x] 2.4 — The documented release gate does not exist (F8)

PORTAINER.md and SECURITY_REPORT.md claim the workflow runs the full isolated suite, Trivy HIGH/CRITICAL image gates, secret scanning and the source preflight before calling Portainer. .gitea/workflows/deploy.yml runs py_compile plus four unit tests, then builds, pushes latest and calls the webhook unconditionally. deploy/production_preflight.py is never invoked — only its unit test runs.

• Either implement the gate or correct both documents. Do not leave the gap.

[x] 2.5 — The preflight has silently decayed (F9)

It blocks by string-matching source. 4 of 6 markers are dead after the R2 refactor: 'This runtime only supports APP_ENV=local', 'Only local S3 storage is supported', "allowed_hosts=['localhost', '127.0.0.1']", "'environment': 'local'". String gates weaken without failing.

• Replace marker matching with behavioural assertions (import the module, assert the adapter classes in use).

[x] 2.12 — Two divergent stack definitions; the docs named the wrong one

Found 2026-09-21 by asking which file Portainer deploys. PORTAINER.md called deploy/stack.yaml "the production stack"; the deployed file is the repository's docker-compose.yml. deploy/stack.yaml came from the first commit and was never deployed — it supplied credentials as Docker secrets where the deployed file uses plain environment variables.

Decision (2026-09-21): keep docker-compose.yml, delete deploy/stack.yaml.

The gain from Docker secrets here is narrower than it sounds. It keeps values out of docker inspect and the Portainer UI, but local/secrets.py loads them into the process environment anyway, and anyone who can read docker inspect is already root or in the docker group and could read the secret files directly. The operator is the only Portainer user, so the main benefit — limiting what a lower-privileged console user can see — does not apply. Maintaining two definitions that drift was the larger real cost.

local/secrets.py stays. It is inert against the deployed file and costs nothing, and it means a stack can switch to Docker secrets later without a code change.

Still open: rotate the R2 secret key. Not because of Portainer, but because it grants read and write over every customer's artwork and has been readable from the stack environment for some time. The operator password is worth rotating with it.

[x] 2.6 — Base images are not pinned (F10)

Dockerfiles default to mutable python:3.12-slim / nginx:1.28-alpine, the workflow passes no digest build-args, and --pull makes builds non-reproducible — while PORTAINER.md documents digest-pinned immutable bases.

[x] 2.7 — pdf.js loaded from CDN without integrity (F11)

Vendored rather than integrity-pinned, so the Site no longer depends on a third party being reachable and honest when a customer opens it. Both files are served from this origin and their provenance is recorded in local/static/vendor/README.md, verified against the SRI digests cdnjs publishes for 3.11.174.

cdnjs.cloudflare.com is gone from script-src, worker-src and connect-src in both gateway templates: scripts and workers are now 'self' plus blob: for the worker the Site builds itself.

Verified in a browser against the running stack: pdf.js loads from /vendor/, the blob worker starts, and a real 7-page PDF parses with no CSP violation. Both browser suites and the full integration suite pass.

Still open: the version. 3.11.174 is old. GHSA-wgrm-67xf-hhpq is mitigated — dtf-site.html already passes isEvalSupported: false, which is the documented workaround — but staying on it indefinitely is not a posture. Upgrading is an API change rather than a file swap and needs its own browser testing, so it is deliberately not bundled here.

[x] 2.8 — Single shared operator credential (F12)

Accounts now live in dtf_local.operators, one per person, with movements.operator and order_files.created_by recording who actually acted. Administered from the API container with python3 -m app.operators (list, add, password, disable, enable); passwords are read from the terminal so they never reach shell history or the process list, and disabling revokes open sessions immediately rather than leaving them valid for the rest of the eight-hour window.

Migration was the risk, since getting it wrong locks the factory out of the Kanban. OPERATOR_EMAIL/OPERATOR_PASSWORD seed the first account, once: a password changed through the CLI is never reverted by a stale environment variable on the next deploy. The first attempt did not work — db-init was not given those variables in either compose file, so no account would have been created and login would have failed closed with 503. Both files now pass them to the migration job. Verified end to end: the unchanged credential still logs in, a second operator authenticates separately, wrong passwords and unknown accounts are rejected, and disabling ends access at once.

Roles are deliberately not included. The meeting described separation of duties for rework authorisation (Mayana classifies, Thales or Alexandre authorise), but the rework feature does not exist in this system, so there is nothing for a role to gate. Building an authorisation model with no consumer would be guesswork. Add roles with the feature that needs them.

[~] 2.9 — TLS is terminated outside the repository (F13)

Downgraded 2026-09-21. The stack publishes plain HTTP on 18080/18081 while COOKIE_SECURE: "true", and nothing in the repo provisions certificates — but nginx-proxy-manager on the host owns 80/443 and terminates TLS in front of it, so cookies are not being dropped in practice. This is undocumented operational knowledge rather than a live defect.

What remains: record the proxy in PORTAINER.md as part of the deployment contract, so nobody moves the stack to a host without one and silently breaks every session cookie. TAREFAS.md A2 still lists the certificate as pending; confirm it is actually issued for the DTF subdomain.

[ ] 2.10 — No email verification, no password recovery (F14)

A locked-out customer has no path back, and registration accepts any CNPJ without proving control of the e-mail. Needs a transactional mail provider — client input.

[ ] 2.11 — LGPD (F15)

CNPJ, phone and e-mail are kept indefinitely in accounts.profile and orders.snapshot. Artwork has a 30-day policy; personal data has none, and there is no privacy notice, consent record or deletion path.

---

Block 3 · Architecture — needs a decision before code

[?] 3.1 — Manual quote approval contradicts the 24h business case (F16)

Payment requires quotes.approved, set only by an authenticated operator. The meeting's premise was that the 17h30 order waiting until 5am is what costs the money. As built, a 2am order still waits for a person. CONTEXT.md frames this as a temporary development trust boundary — the risk is that it silently becomes the delivered model.

Decide: what makes a quote auto-approvable (mode, metre range, grade floor, returning customer), and what still routes to a human.

[?] 3.2 — Billable metres are computed in the customer's browser (F17, F18)

For loose artwork, metros comes from desenhaMontagem/encaixar — a canvas alpha-mask packer running client-side. The server never recomputes it. passoDe()/CELULAS_MAX coarsen the grid for large sheets and image decoding differs by browser, so the same cart can price differently on different devices. The code comments reference "o motor do servidor"; that engine does not exist.

Worse, the layout the customer is quoted on is never produced — operators upload final files by hand, so billed metres ≠ printed metres and a designer redoes work the site already did.

Decide: port the packer to the server as the pricing authority and the print-file generator, with the browser as preview only. This is the single largest gap between what was promised in the meeting and what exists.

[?] 3.3 — The 5 GB problem is unsolved (F19)

Transport accepts 5 GiB; SCAN_MAX_BYTES / ClamAV StreamMaxLength release only ≤ 128 MiB. Files above that are quarantined permanently with no path forward. This is exactly the risk Jorge raised in the meeting.

Decide: raise the scan ceiling with a resource/timeout design, or define an explicit large-file path (staged scan, sampled scan, operator override with audit).

[ ] 3.4 — Upload throughput (F20)

8 MiB parts, strictly sequential in local/static/upload.js:21, one presign round-trip per part → ~640 sequential API calls for a 5 GB file, through an nginx limit_req of 20r/s. Add parallelism (4–6 in flight) and batch presigning.

[ ] 3.5 — Payment ordering (F27)

dev_paid charges before persisting the order and passes no idempotency key. Harmless with FakePayment; with Mercado Pago that ordering is how you get double charges. Fix as part of 1.1.

---

Block 4 · Scale and performance

• [x] 4.1 — Ten indexes added, each matched to a query the application issues, and no more: every extra index is paid for on each write. The outbox and live uploads use partial indexes so they stay the size of the backlog rather than of all history. Verified against the running database — the planner chooses outbox_pending and orders_owner for the queries they exist for.
• [x] 4.2 — The board returns every order still in progress, however old, plus a window of recent finished ones (BOARD_FINISHED_LIMIT, default 50) and the true finished total. An operator can never lose a card they could act on; only terminal ones are trimmed. The Kanban column reads "Finalizado · 50 de 213" when truncated, so the count is not mistaken for an all-time total. Pending quotes are capped too.
• [ ] 4.3 — Scan throughput (F21): one scan_loop thread, worker at replicas: 1, ClamAV MaxThreads 2, browser gives up after 150s.
• [ ] 4.4 — Quality grade fallback (F22): when carregarImagem fails, px(f)=Math.sqrt(f.size/1024)*95 stands — a DPI inferred from file size in bytes — and it drives up to a 25% discount. Fail closed instead.
• [x] 4.5 — Dead config (F28): resolved by deleting deploy/stack.yaml in 2.12. CLAMD_HOST no longer appears anywhere; scanning.py reaching 'scanner' directly is now simply how it works, not a contradiction.

---

Block 5 · Hygiene and maintenance

• [x] 5.1 — CI coverage (F33). An integration job now builds the localhost stack and runs smoke, workflow, security, scanning, retention, runtime security and both browser suites; publish-and-deploy depends on it. Verified by reintroducing the 0.1 defect: py_compile and the unit tests still passed while smoke_test failed on /session, blocking the release. Browser tests skip with a warning when the runner has no Chrome — install google-chrome-stable on the runner (or set CHROME_BIN) to make them gate as well.

• [x] 5.2 — portal/, kanban/, agente/, the root schema.sql and .env.exemplo removed: 2,283 lines implementing a model this system abandoned, referenced by nothing, with several endpoints taking the acting user from the request body. The documents describing them are archived under docs/historico/ with a header saying they are background, not instructions.

• [x] 5.3 — Root requirements.txt deleted. It pinned by wildcard, listed packages the system does not use, and sat next to the hash-locked infra/requirements.lock inviting the wrong one to be installed. Only historical documentation referred to it.

• [x] 5.4 — pip removed from infra/requirements.txt and from the lock. Nothing depended on it; it was pinned only because it was listed directly, and installing it put a package manager inside the read-only runtime image. The base image's own pip performs the hash-enforced install. Verified: the image builds under --require-hashes and reports the base pip, 25.0.1.

• [ ] 5.5 — Doc drift (F34). README.md, CONTEXT.md, LOCAL_SETUP.md and SECURITY_REPORT.md describe a MinIO localhost stack, an API with "no external network route", a operator / local-operator-only login the email-validated model rejects, and a release gate — none match the current tree.

• [x] 5.6 — The Site's 1,575-line inline script is now nine files under local/static/, cut at the author's own section boundaries so no function was split: config, modes, upload, sheet analysis, PDF, quality, packing, cart, flow. dtf-site.html is 1,394 lines of markup and style. The extraction was verified byte-identical before the tags were swapped in, and they load as classic scripts in the original order, so evaluation semantics are unchanged.

  A consequence worth having: with no inline script anywhere, the policy needs no hash allowlist and is now simply script-src 'self'. site-packing.js is also where a server-side packer (3.2) has to agree, which was the point of splitting.

• [ ] 5.7 — Commercial rules duplicated between FAIXAS (JS) and TIERS (Python) (F26). test_pricing guards parity; generate one from the other instead.

• [ ] 5.11 — Nothing tests the schema against an empty database. The 4.1 indexes were added next to the existing one, which sits before the tables they name, so CREATE INDEX ... ON dtf_local.order_files ran before that table existed. Every local run passed because the volume already had the tables; CI caught it on its clean volume. A migration is only really exercised from nothing, so the integration job should run down -v before up — or a dedicated step should apply schema.sql twice to a fresh database, proving both a first install and a re-run.

• [ ] 5.10 — The browser suites do not run in CI. Chrome runs in the runner container and can only reach the stack through ports published on the host, which is a different network namespace when the runner is itself a container. The API, workflow, security, scanning, retention and runtime suites were moved inside the stack's network and do gate. The browser suites are the only coverage for the artwork editor and the full customer journey, so they need either Chrome in a container on that network, or a runner with host networking. Until then they gate locally only, and CI warns when it skips them.

• [ ] 5.9 — local/browser_test.mjs failed once and passed on an immediate re-run, with no code change in between (2026-09-21). It is a deploy gate when the runner has Chrome, so an intermittent failure there blocks releases for no reason. Suspect Chrome startup timing or a race against stack readiness. Watch it, and if it recurs add an explicit readiness wait rather than a retry.

• [x] 5.8 — The Site claimed 90-day file storage and 12-month history in two places, and invited customers to reorder "sem subir de novo". Files are kept 30 days. The copy now states 30 days, says a later order needs the file again, and keeps only the true part: order history remains in the account. Policy unchanged; the promise was corrected to match it.

---

Done

Week 1 — infrastructure, uploads, service base

• [x] Compose stack: Site, Kanban, API, PostgreSQL, worker, ClamAV, R2/MinIO storage.
• [x] Gitea + Portainer publication path; web service startup and API rollout fixed.
• [x] Database passwords with special characters handled via discrete libpq fields.
• [x] Kanban e-mail/password login; missing operator config no longer breaks stack boot.
• [x] Direct resumable multipart browser → private object storage.
• [x] Quarantine + ClamAV gate: only clean files can be quoted, paid, downloaded or queued.
• [x] Retention worker: incomplete 1d, rejected 3d, originals 7d after approval, finals 30d.
• [x] Server-side pricing authority with parity test against the Site JavaScript (4,444 cases).
• [x] Loose-artwork packing flow: per-file width, copies, rotate, mirror, live 57 cm preview, 5 mm gap, ruler and watermark preserved; metres follow packed height.
• [x] Rotation/mirror applied to packing masks; stale renders no longer overwrite a newer preview.
• [x] Hash-locked Python dependencies; Trivy reports in output/security/.

Block 0 — 2026-09-18

• [x] import os restored in local/auth.py; COOKIE_SECURE is now a single module constant shared with local/app.py.
• [x] Navigation links no longer preselect a product (data-modo-cta removed).
• [x] Ready sheet vs loose artwork is an explicit, priced, reversible selector (#tipoEnvio), shown in all four modes and locked once a file is attached. The product is the declaration; sel() no longer switches anything silently.
• [x] medirFolha returns dpiFolha; an image that cannot span the film width at DPI_RECUSA is refused as a sheet, with one click to send it as loose artwork.
• [x] pintaCaminhos scoped to #caminhos .cam — its global .cam selector was clobbering the new control.
• [x] compose.local.yaml restored (MinIO, fake providers, builds from source).
• [x] APP_DB_PASSWORD separated from POSTGRES_PASSWORD, with a bootstrap.py guard that refuses identical credentials in both configuration forms.
• [x] Checkout confirmation moved out of the panel the success path hides.

Block 2 and CI — 2026-09-18

• [x] The web gateway overwrites X-Forwarded-For with the peer address instead of appending to it, and client_ip() resolves the requester for rate-limit buckets and security events. Verified: a forged 203.0.113.99 never reaches the audit trail.
• [x] Guest sessions are limited per source. The first attempt used 30/IP, which the new regression caught as too tight for shared NAT — recreating the original fault in a narrower form — so the ceiling is 240 per 15 minutes, overridable with GUEST_SESSION_LIMIT.
• [x] Security events now carry the source address (operator_login_failed, customer_login_failed, cross_origin_rejected, http_security_event).
• [x] CI runs the integration suites against a real stack before publishing.

2026-09-21

• [x] 2.3 — local/secrets.py resolves every <NAME>_FILE into <NAME> from the API, worker and bootstrap entrypoints, failing closed on an unreadable or empty secret and on a value supplied both ways. Verified by booting the API, the worker and bootstrap with credentials supplied only as mounted files, including a password containing :/?#[]&=+$ ,%. OPERATOR_USER in the stack became OPERATOR_EMAIL, which is what the runtime reads.

• [x] 2.5 — The gate now loads local/secrets.py and makes it resolve every secret deploy/stack.yaml declares, plus asserts it fails closed. Verified against a no-op loader (11 blockers) and one that swallows a missing file (1 blocker); only the real implementation passes. The four marker strings that stopped matching when R2 support landed were removed; the two describing real blockers stay, so the gate still refuses a release while the payment and messaging adapters are fake.

• [x] 2.4 — A blocking Trivy secret scan was added and verified both ways: a planted AWS key pair, GitHub token and private key block the job; the repository passes clean. Worth knowing: Trivy allowlists documented example credentials, so my first probe passed with AWS's own sample keys — the gate is a backstop, not permission to commit secrets. The source preflight now runs and always prints its verdict, enforcing only when ENFORCE_PRODUCTION_PREFLIGHT is true; enforcing it today would block every deploy, since it refuses a release while the adapters are fake. Image vulnerabilities are reported after each build, not enforced — 56 HIGH and 3 CRITICAL, only 15 with an upstream fix. PORTAINER.md and SECURITY_REPORT.md now carry a table of what gates and what does not, replacing descriptions of checks that never ran.

• [x] 2.6 — Both bases pinned by digest, OS packages upgraded in the production images, and the web image moved off the nginx 1.28 line.

  Image	Before	After
  API	56 HIGH, 3 CRITICAL (15 fixable)	46 HIGH, 0 CRITICAL
  Web	5 HIGH, all unfixable in place	0 HIGH, 0 CRITICAL

  The 1.28 nginx pins nginx=1.28.3-r1 in /etc/apk/world, so apk upgrade cannot patch it even though Alpine ships -r7; nginx:alpine (1.31.6) is clean while 1.29-alpine scans worse at 37 HIGH. The two remaining "fixable" API findings are msgpack and setuptools, which I confirmed are absent from the built image rather than trusting the earlier report. Local images now share the pinned bases, so the integration suite exercises what ships; full suite passes on nginx 1.31.6. With both images at zero CRITICAL, the image scan now gates on CRITICAL and reports HIGH.

2026-09-21 — from the runner host inventory

• [x] Fixed a regression in 2.1: the production gateway sits behind nginx-proxy-manager, so $remote_addr there is the proxy, not the customer. Overwriting X-Forwarded-For with it would have recorded the proxy's address for every request in production — the same bug 2.1 set out to fix. The gateway now uses real_ip to recover the customer's address from the proxy's header, trusting only private networks, so a request arriving directly at the published port cannot spoof it. Validated with nginx -t against the rendered config.
• [~] 2.9 downgraded: TLS is terminated by that proxy, not missing.

Reporting

• [x] Week-1 client report (Relatorio-Semana-1-DTF.docx), corrected 2026-09-18 to remove the inaccurate "Arquivo por metro permanece separado, com seleção explícita" claim and the internal commit reference.